Pith. sign in

REVIEW 3 major objections 6 minor 21 references

High-Performance Reinforcement-Learned BP Decoding of Quantum LDPC Codes

T0 review · 3 major / 6 minor · reviewed 2026-07-31 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A second-order local BP update, driven by a learned schedule, beats heavy BP-OSD on quantum LDPC codes while staying purely local.

desk verdict Solid empirical decoder paper: second-order local cascades on a frozen RL-S schedule beat BP-OSD-10 at far fewer sweeps, but no ablation shows the learned table is doing real work. read the letter →

arxiv 2607.24891 v1 pith:MXT7PUWR submitted 2026-07-27 cs.IT math.IT

Mohsen Moradi , Vahid Nourozi , Taejoon Kim , Remi A. Chou , David G. M. Mitchell This is my paper · ORCID
classification cs.ITmath.IT MSC 94B3581P73
keywords quantumLDPCcodesbeliefpropagationreinforcementlearningsequentialschedulingsecond-orderlocalupdateCSSdepolarizingchannelBP-OSD
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Quantum LDPC codes need fast, local decoders, but ordinary belief propagation often stalls because of short cycles and many errors that look the same on the syndrome. Earlier work showed that a reinforcement-learned order for updating variable nodes can help BP converge. This paper keeps that same offline-learned schedule table and, at inference time, expands each chosen update into a small cascade: the selected node plus its neighbors that share a check. The result, RL-S2LU, remains a purely local message-passing decoder with bounded degree cost, yet on several finite-length quantum LDPC codes it corrects more errors with far fewer sweeps than flooding BP or BP with order-10 OSD post-processing. A sympathetic reader cares because high-threshold quantum memory designs need decoders that are both accurate and hardware-local; this shows schedule learning plus a cheap neighborhood cascade can close much of the gap without algebraic post-processing or retraining.

What carries the argument

The second-order neighborhood U2(v): after the learned policy picks an anchor v, the decoder runs the quaternary sequential VN update on an ordered traversal of v and all variables that share at least one adjacent stabilizer check with v, while only the anchor is removed from the without-replacement schedule set.

What would settle it

Retrain or ablate the Q-table under the actual second-order residual dynamics on the [[288,12,18]] code; if the reported block-error gains over BP-OSD-10 at T=10 vanish or reverse once the schedule is no longer the unmatched RL-S table, the central transfer claim fails.

Watch

Extended reading notes

Core claim

RL-S2LU reuses the trained RL-S Q-table and greedy anchor rule, but replaces each selected variable-node update with a deterministic second-order local cascade over that node and the variable nodes sharing an adjacent X- or Z-check. On the tested bivariate-bicycle and A5 codes, this purely BP decoder with small sweep caps outperforms conventional flooding BP and BP-OSD-10 run with far more iterations, and typically converges within the first outer sweep in the low-error regime.

Load-bearing premise

The offline-trained schedule table still picks good anchors when each action suddenly updates a whole local neighborhood and changes residual states in ways the original training never saw.

Editorial extensions

If this is right

  • Learned sequential BP for QLDPC can be strengthened by local cascade propagation without retraining or OSD.
  • On the [[288,12,18]] BB code, T=10 RL-S2LU sweeps already beat BP and BP-OSD-10 at T=1000 over the simulated depolarizing range.
  • Raising the sweep cap further cuts non-convergence failures and improves the low-p block error rate while average outer sweeps stay near one in the low-error regime.
  • For bounded-degree QLDPC families the extra work stays O(Tn) with a larger constant, and the cascade is locally parallelizable.
  • On shorter codes the remaining failures become mostly logical rather than non-convergence, pointing to a logical-error-limited regime under pure BP.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the unmatched Q-table already works this well, a schedule trained explicitly on second-order residual transitions might unlock still lower error floors or fewer sweeps.
  • The same cascade-after-learned-anchor idea may transfer to other sparse quantum or classical codes where flooding stalls on short cycles, without needing OSD hardware.
  • Hardware implementations could treat each anchor’s U2 neighborhood as a small parallel tile, trading local wiring for fewer global iterations in quantum memory controllers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The manuscript proposes RL-S2LU, a modification of the authors' RL-S sequential BP decoder for CSS QLDPC codes. RL-S uses a Q-table trained offline over residual-syndrome states (Eqs. (8)-(10)) to greedily select anchor variable nodes; RL-S2LU keeps the same frozen table and anchor rule but replaces the single-VN update with a deterministic second-order cascade: after an anchor v* is selected, all VNs sharing a check with v* (the set U2(v*), Eq. (11), of size at most Umax = 1 + ∆v(∆c-1)) are updated in a fixed graph order (Algorithm 1). Simulations on the [[288,12,18]] and [[144,12,12]] bivariate-bicycle codes and the [[180,10]] A5 code show that RL-S2LU with sweep cap T=10 achieves lower block error rate than flooding BP and BP-OSD-10 at T=1000 over the simulated depolarizing range, with further low-p gains at T=100/1000, and typically converges within the first outer sweep (Table II). The message equations (1)-(7), residual state, algorithm, and complexity bounds are clearly specified.

Significance. If the results hold, the paper demonstrates a practically useful fact: a schedule learned under single-VN-update dynamics can be combined at inference time, without retraining, with a deterministic local propagation cascade, yielding large BLER gains while remaining purely local BP (no OSD post-processing). The presentation is a strength: Eqs. (1)-(11) and Algorithm 1 are complete enough to reimplement, the convergence/termination criterion is explicit, and the failure decomposition in Table I (non-convergence vs. logical errors) is informative. The parameter count is small (frozen Q-table, sweep cap T, channel initialization), and the claims are concrete and falsifiable. The contribution is incremental within the authors' RL-scheduling line [15],[18]-[21], and its meaning depends on how much of the gain is attributable to the learned schedule versus the cascade (see Major Comment 1); even in the worst case, the empirical decoder itself appears strong.

major comments (3)
  1. [§III.C, §IV, Figs. 1-3] §III.C / §IV, Figs. 1-3: there is no ablation isolating the contribution of the learned schedule from that of the deterministic U2 cascade. The Q-table QRL was trained in [19] under single-VN-update dynamics, where each anchor action changes one node's beliefs and residuals; here each action triggers up to Umax = 1+∆v(∆c-1) ~ 31 VN updates for the BB codes (§III.D), so the state-transition structure the table was optimized for is substantially altered. Two readings are consistent with the presented curves: (a) the learned ranking remains informative under the cascade, or (b) the cascade does essentially all the work and almost any anchor order (random, or a simple residual-weight heuristic) paired with the same U2 update would match the reported BLER. The paper's framing — 'learned sequential BP schedules can be strengthened by local propagation without retraining' (§V) — and its name re
  2. [§I, §IV, Fig. 1, Table II] §I, §IV, Fig. 1: the headline comparison 'RL-S2LU with only T=10 decoder sweeps outperforms BP and BP-OSD-10 with T=1000 iterations' conflates unlike units of work. One RL-S2LU sweep performs up to n·Umax local VN updates (§III.D), so T=10 sweeps cost up to ~10·288·31 ≈ 9×10^4 VN updates on the [[288,12,18]] code, versus 1000·288 ≈ 2.9×10^5 VN updates for flooding BP at T=1000 (plus the uncounted order-10 OSD Gaussian-elimination overhead). The claim appears to survive honest accounting, but the paper should state it in terms of measured average local-update counts (or wall-clock equivalents) per decoding, alongside Table II's sweep counts, and note that RL-S2LU's serial latency is worse than flooding BP's by the Umax factor under a fully serial implementation (§III.D acknowledges this only asymptotically).
  3. [§IV, Fig. 1, Table I] §IV: no statistical methodology is reported. Fig. 1 shows BLER down to 10^-7 at p=0.05 and claims 'about one order of magnitude gain over BP-OSD-10 at small values of p,' yet the number of simulated error events/samples per point and confidence intervals are nowhere given, and Table I reports percentages (e.g., 27.0% non-convergence at p=0.05, T=1000) conditioned on error events that must be rare at those BLERs. The simulation protocol (sample counts per p, stopping rule, whether OSD-10 failures were counted identically) should be stated and error bars or sample sizes added to Figs. 1-3 and Table I.
minor comments (6)
  1. [§III.C, Eq. (11)] Eq. (11): U2(v) as defined already contains v (since v ∈ ∂c for every c ∈ ∂v), so the sentence 'U2(v) contains v and all VNs that share at least one adjacent check' is redundant; also, 'second-order' usually denotes distance-2 in graph nomenclature, whereas U2 is the closed distance-2 neighborhood — a clarifying sentence would help.
  2. [Figs. 1 and 3] Fig. 1 legend lists RL-S at T=10 but no such curve is discernible, and Fig. 3 omits the RL-S2LU T=100/1000 curves shown in Figs. 1-2, making the cross-code comparison inconsistent. Please make the plotted configurations uniform across the three figures or explain the omissions.
  3. [Table II] Table II is titled 'average number of outer iterations' while the text uses 'sweeps'; one row (RL-S2LU, T=1000, p=0.08, value 1.3) jumps to 3.2 at p=0.09 and 9.6 at p=0.10, which is worth one sentence of interpretation, since it bears on the latency claims in §III.D.
  4. [§II.A, Eq. (2)] §II.A: the initialization µ = log((1-2p/3)/(2p/3)) assumes the decoder knows the true physical error rate p; please state whether any p-mismatch sensitivity was tested, and note that κP in Eq. (2) embeds the same assumption.
  5. [§IV] Comparison set: BP-OSD-10 is the only strong baseline; a brief note positioning RL-S2LU against other purely-BP improvements (e.g., normalized/offset min-sum variants for degenerate codes, or the cluster-based sequential decoder of the companion work [21]) would clarify where the claimed gain is genuinely new versus inherited from sequential scheduling generally [15].
  6. [Title page, References] Typesetting artifacts in the author block ('Ari zona', 'Universit y', 'V ahid Nourozi') and a few missing-space artifacts ('VNs', 'U2(v)') should be cleaned up; [20],[21] are cited as arXiv-only 2026 preprints — if [19] is still unpublished, the reproducibility of the inherited Q-table should be addressed (see confidential remarks).

Circularity Check

1 steps flagged · score 1.0 of 10

No derivation circularity: empirical decoder gains measured against external BP/BP-OSD baselines; only mild non-load-bearing reuse of authors' prior RL-S table.

  1. self citation load bearing [§III opening; §III.B eq. (10); Abstract]
    "RL-S2LU is an inference-time second-order local-update extension of RL-S [19]. It uses the same trained table QRL, residual state, and greedy anchor-selection rule as RL-S. ... v⋆ = arg max_{v∈R} QRL(σ(v), v)"

    The learned component of the named decoder is entirely inherited from the authors' prior RL-S paper without retraining or independent re-derivation. This is mild and not load-bearing for the empirical BER claims (which rest on external BP/BP-OSD comparisons), but the schedule itself is justified only by that self-citation chain rather than by a self-contained argument in this manuscript.

full rationale

This is an empirical systems paper, not a first-principles derivation. The central claims are measured block-error-rate curves (Figs. 1–3, Tables I–II) for RL-S2LU versus conventional flooding BP and BP-OSD-10 on three published QLDPC codes. Those baselines are independent of the authors' RL machinery, so the reported gains are not forced by construction from fitted inputs or self-defined quantities. The method reuses the offline QRL table, residual state, quaternary VN update, and greedy anchor rule from the authors' prior RL-S work [19], and extends inference with a deterministic second-order neighborhood cascade U2(v). That is ordinary incremental self-citation of a building block, not a circular reduction: nothing in the paper equates a 'predicted' BER to a quantity already fitted into QRL, and no uniqueness theorem is imported to forbid alternatives. The open scientific question (whether gains come from the learned schedule versus the cascade alone) is an ablation/novelty issue, not circularity under the stated criteria. Score 1 only to mark the non-load-bearing dependence on [19] for the schedule component.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

Load-bearing content is algorithmic and empirical. The claim rests on standard CSS/depolarizing modeling, the inherited RL-S message and state machinery, the unretrained Q-table, hand-chosen sweep caps T, and the definition of the second-order neighborhood action. No new physical entity is postulated; the invented piece is the decoder rule itself.

free parameters (4)
  • sweep cap T = 10, 100, 1000 (reported settings)
    Outer-iteration budget chosen for experiments (10/100/1000); central comparisons hinge on small-T RL-S2LU vs large-T baselines.
  • trained QRL table = inherited from RL-S [19], values not given
    Offline RL values fully determine greedy anchor selection; not retrained for S2LU and not published here.
  • channel LLR init μ and Pauli score offsets κP = μ=log((1-2p/3)/(2p/3)); κ from p as in §III-A
    Fixed from depolarizing p via stated formulas; standard but p-dependent initialization that affects all BP variants compared.
  • OSD order = 10 = 10
    Baseline post-processing strength chosen by authors; affects the claimed gain margin versus BP-OSD-10.
assumptions (5)
  • domain assumption CSS QLDPC codes with HX HZ^T = 0 and independent depolarizing noise correctly model the simulated channel and syndromes.
    §II-A defines the entire evaluation setting; no coherent noise, circuit-level noise, or measurement error.
  • domain assumption Quaternary sequential VN BP updates (1)–(7) and residual-syndrome state (8)–(9) are valid local BP machinery for these codes.
    Inherited wholesale from RL-S [19] in §III-A–B; correctness of gains assumes this BP core is implemented as specified.
  • ad hoc to paper Greedy use of a fixed offline QRL table yields a meaningful adaptive schedule under the altered S2LU residual dynamics.
    §III explicitly reuses QRL without retraining; this is the key unproven transfer assumption behind the method.
  • domain assumption Block error rate counting non-convergence and logical error is the right success metric; convergence means matching both CSS syndromes.
    §II-A and §IV; standard in the subfield but shapes Table I interpretation.
  • standard math For bounded-degree QLDPC families, |U2(v)| is O(1), so work remains O(Tn) up to constants.
    §III-D complexity argument from graph degree bounds.
invented entities (1)
  • RL-S2LU decoder (second-order local-update cascade U2/L2 after RL anchor selection)
    purpose: Propagate each learned scheduling decision farther on the Tanner graph while keeping updates local and avoiding OSD post-processing.
    Defined in §III-C and Algorithm 1; the paper’s primary constructed object. Empirical handle only via the reported simulations, not an external physical prediction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of High-Performance Reinforcement-Learned BP Decoding of Quantum LDPC Codes." pith.science (2026). https://pith.science/paper/MXT7PUWR

@misc{pith2026260724891,
  author       = {Pith},
  title        = {Pith review of: High-Performance Reinforcement-Learned BP Decoding of Quantum LDPC Codes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MXT7PUWR}},
  note         = {Machine review of arXiv:2607.24891}
}
read the original abstract

Belief-propagation (BP) decoding is attractive for quantum low-density parity-check (QLDPC) codes because it uses local message passing on sparse Tanner graphs. However, conventional flooding BP often stalls due to stabilizer degeneracy and short cycles. Reinforcement-learning-based sequential variable-node scheduling (RL-S), which learns the update order offline, has shown that adaptive scheduling can improve BP convergence. In this paper, we extend this idea with a second-order local update decoder, RL-S2LU. The proposed decoder preserves BP locality and low complexity, while numerical results show significant error-correction gains over conventional BP and the considered BP-OSD-10 baseline.

Figures

Figures reproduced from arXiv: 2607.24891 by the authors.

Figure 1
Figure 1. Error-correction performance comparison for the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Error-correction performance comparison for the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Error-correction performance comparison for the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 5 linked inside Pith

  1. [15]

    S equential BP-based decoding of QLDPC codes,

    M. Moradi, S. Habib, V . Nourozi, and D. G. M. Mitchell, “S equential BP-based decoding of QLDPC codes,” in ICC 2026 - IEEE International Conference on Communications , 2026, pp. 1–6

    2026

  2. [18]

    Enhancing be lief prop- agation decoding of polar codes: A reinforcement learning a pproach,

    M. Moradi, S. Habib, and D. G. M. Mitchell, “Enhancing be lief prop- agation decoding of polar codes: A reinforcement learning a pproach,” IEEE Communications Letters , vol. 29, no. 6, pp. 1285–1289, 2025

    2025

  3. [21]

    Learning to decode quantum LDPC codes via cluster- based sequential belief propagation,

    M. Moradi, T. Kim, R. A. Chou, and D. G. M. Mitchell, “Learning to decode quantum LDPC codes via cluster- based sequential belief propagation,” 2026. [Online]. Ava ilable: https://arxiv.org/abs/2607.20130

    arXiv 2026

  4. [19]

    L earning to decode quantum LDPC codes via belief propagation,

    M. Moradi, V . Nourozi, S. Habib, and D. G. M. Mitchell, “L earning to decode quantum LDPC codes via belief propagation,” arXiv preprint arXiv:2603.10192, 2026

    arXiv 2026

  5. [1]

    Scheme for reducing decoherence in quantum c omputer memory,

    P . W. Shor, “Scheme for reducing decoherence in quantum c omputer memory,” Physical Review A , vol. 52, no. 4, pp. R2493–R2496, 1995

    1995

  6. [2]

    Stabilizer codes and quantum error corre ction,

    D. Gottesman, “Stabilizer codes and quantum error corre ction,” Ph.D. dissertation, California Institute of Technology, 1997, a rXiv:quant- ph/9705052

    arXiv 1997

  7. [3]

    Quantum LDPC codes with posi tive rate and minimum distance proportional to the square root of the bloc klength,

    J.-P . Tillich and G. Zémor, “Quantum LDPC codes with posi tive rate and minimum distance proportional to the square root of the bloc klength,” IEEE Transactions on Information Theory , vol. 60, no. 2, pp. 1193– 1202, 2014

    2014

  8. [4]

    Quantum Kronecker sum-p roduct low-density parity-check codes with finite rate,

    A. A. Kovalev and L. P . Pryadko, “Quantum Kronecker sum-p roduct low-density parity-check codes with finite rate,” Physical Review A , vol. 88, no. 1, p. 012311, 2013

    2013

Show all 21 references
  1. [5]

    Asymptotically good quan tum and locally testable classical LDPC codes,

    P . Panteleev and G. Kalachev, “Asymptotically good quan tum and locally testable classical LDPC codes,” in Proceedings of the 54th Annual ACM SIGACT Symposium on Theory of Computing , 2022, pp. 375–388

    2022

  2. [6]

    Quantum Tanner codes,

    A. Leverrier and G. Zémor, “Quantum Tanner codes,” in Proceedings of the 63rd IEEE Annual Symposium on F oundations of Computer Sc ience, 2022, pp. 872–883

    2022

  3. [7]

    High-threshold and low-overhead fault-toler ant quantum memory,

    S. Bravyi, A. W. Cross, J. M. Gambetta, D. Maslov, P . Rall, and T. J. Y oder, “High-threshold and low-overhead fault-toler ant quantum memory,” Nature, vol. 627, pp. 778–782, 2024

    2024

  4. [8]

    Low-density parity-check codes,

    R. G. Gallager, “Low-density parity-check codes,” IRE Transactions on Information Theory , vol. 8, no. 1, pp. 21–28, 1962

    1962

  5. [9]

    Good error-correcting cod es based on very sparse matrices,

    D. J. C. MacKay and R. M. Neal, “Good error-correcting cod es based on very sparse matrices,” IEEE Transactions on Information Theory , vol. 45, no. 2, pp. 399–431, 1999

    1999

  6. [10]

    On the iterative decoding of spa rse quantum codes,

    D. Poulin and Y . Chung, “On the iterative decoding of spa rse quantum codes,” Quantum Information and Computation , vol. 8, no. 10, pp. 987– 1000, 2008, arXiv:0801.1241

    2008 arXiv

  7. [11]

    Degenerate quantum LDPC codes with good finite length performance,

    P . Panteleev and G. Kalachev, “Degenerate quantum LDPC codes with good finite length performance,” Quantum, vol. 5, p. 585, 2021

    2021

  8. [12]

    Dec oding across the quantum low-density parity-check code landscape,

    J. Roffe, D. R. White, S. Burton, and E. T. Campbell, “Dec oding across the quantum low-density parity-check code landscape,” Physical Review Research, vol. 2, no. 4, p. 043423, 2020

    2020

  9. [13]

    Shuffled belief propaga tion decoding,

    J. Zhang and M. P . C. Fossorier, “Shuffled belief propaga tion decoding,” IEEE Transactions on Communications , vol. 53, no. 2, pp. 209–213, 2005

    2005

  10. [14]

    A reduced complexity decoder architect ure via layered decoding of LDPC codes,

    D. E. Hocevar, “A reduced complexity decoder architect ure via layered decoding of LDPC codes,” in IEEE W orkshop on Signal Processing Systems, 2004, pp. 107–112

    2004

  11. [16]

    Reinforcement learning for sequential decoding of generalized LDPC codes,

    S. Habib and D. G. M. Mitchell, “Reinforcement learning for sequential decoding of generalized LDPC codes,” arXiv preprint arXiv:2307.13905, 2023

    2023 arXiv

  12. [17]

    RELDEC: Reinforce ment learning-based decoding of moderate length LDPC codes,

    S. Habib, A. Beemer, and J. Kliewer, “RELDEC: Reinforce ment learning-based decoding of moderate length LDPC codes,” arXiv preprint arXiv:2112.13934, 2021

    2021 arXiv

  13. [20]

    Learning-based list s equential belief propagation decoding of quantum LDPC codes,

    M. Moradi, T. Kim, and R. A. Chou, “Learning-based list s equential belief propagation decoding of quantum LDPC codes,” 2026. [ Online]. Available: https://arxiv.org/abs/2606.20926

    2026 arXiv

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.