acceptodds
Under review as a conference paper at ICLR 2027

PRIME: A Neural-Prior BP-OSD Decoder for Quantum LDPC Codes

Abstract

Quantum error correction protects logical information by encoding it redundantly across physical qubits, and a decoder infers a correction from the measured syndrome without introducing a logical error. For bivariate bicycle (BB) quantum low-density parity-check (qLDPC) codes, belief propagation with ordered-statistics decoding (BP-OSD) is widely used, and its channel prior, the assumed per-qubit error probability, comes from a predefined noise model. We introduce PRIME, which augments BP-OSD with a graph neural network (GNN) that maps each measured syndrome to a syndrome-conditioned channel prior over physical-qubit errors. The underlying BP-OSD procedure is used unchanged, which preserves syndrome consistency while letting the decoder exploit noise structure learned from data, and because the GNN operates directly on the Tanner graph, one model trained on small BB codes generalizes without retraining to larger BB codes and toric codes. We further introduce Jitter-PRIME, a perturbed-prior ensemble whose candidates are aggregated into logical equivalence classes, together with a learned confidence gate that sends only the riskiest instances to it. At physical error rate \(p=0.10\), Jitter-PRIME reduces logical error rate (LER) over BP-OSD-CS-6 by \(6.6\times\), \(9.0\times\), and \(57.8\times\) on BB codes with 288, 360, and 756 qubits. On the two held-out codes, the gate keeps the full ensemble gain with about 15% of the always-on ensemble's solver calls. Under circuit-level noise, Jitter-PRIME reduces LER over BP-OSD-CS-6 by \(1.5\)–\(15\times\) on the length-72 and length-144 BB codes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Graph Neural Networks, Quantum Error Correction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rpVjOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PRIME uses a GNN to predict error probabilities from the measured syndromes, then passes these probabilities to an unchanged BP-OSD decoder. Jitter-PRIME adds several decodes with perturbed probabilities and combines the candidates by logical class. A learned gate selects which samples receive this extra computation. The paper evaluates transfer from small training codes to larger BB codes and to toric codes.

Critical strengths

Learning the inputs to an existing decoder is a useful approach. The single-decode results already improve over BP-OSD, so the gains do not all come from running more solver calls.

Using the same prior network on larger codes is a useful result. The gate also has a helpful control: it performs much better than random routing with the same average call budget.

Critical weaknesses

The circuit-level experiments use a different method from the main neural decoder. Section 4.6 perturbs native fault probabilities without using the GNN, and also changes candidate selection. These results support the perturbation ensemble, but do not show that the learned prior works under circuit-level noise.

Questions for authors

For the circuit-level results, how many baseline successes were tested with the ensemble? Could you provide confidence intervals that include uncertainty from this sampling procedure?

The gate selects the top 10% of samples within a batch. How well does a fixed threshold, chosen on separate validation data, retain the accuracy and cost benefits on the held-out codes?

1 comment
AuthorsOpenReview ↗

Response to the Review

We thank the reviewer for the careful reading and for two well-targeted questions. Following the reviewer's suggestions, we ran both controls, and they strengthen the circuit-level results. Section, table and appendix numbers refer to the revised PDF.

W1. Circuit-level results and the learned prior

We agree with the reviewer's reading. The circuit-level experiments evaluate the perturbation ensemble and gate with native priors, and the revision makes this explicit in the section title ("Circuit-Level Decoding with Native Priors", Section 5.3) and the abstract.

The ensemble and gate are strong at circuit level: at 10% routing, the gate finds 79% of BP-OSD failures, against 30% for syndrome-weight ranking and 19% for BP non-convergence.

We also tested the GNN prior itself in the circuit-level memory experiments at p=.005, in four ways on n=144 (Appendix D.7, Table 7):

  1. A GNN trained on the detector error model, predicting a bounded correction to each fault mechanism's native prior: 3.02% vs. 3.00% LER (learned vs. native).
  2. That GNN as an extra candidate next to the native decode, against an extra native-jitter candidate at equal calls: 2.20% vs. 2.30%.
  3. Our shipped code-capacity checkpoint, unchanged, as a per-round prior on the fault mechanisms that act as single-qubit errors: 3.40% vs. 3.46%.
  4. The same checkpoint as features for the routing gate, against matched adjacent-fired-check features: recall at 10% routing changes by −1.3 points, CI [−3.7, +1.6].

In every case the learned prior performs on par with the native prior. At circuit level the syndrome already locates most faults (simple adjacent-check counting reaches AUROC .95), which leaves little room for any prior. This confirms the reviewer's reading: the circuit-level gains come from the ensemble and gate.

Q1. Baseline successes tested, and intervals that include this sampling

Following the reviewer's suggestion, we audited every circuit-level setting (Appendix D.6, Table 6). Across all settings we ran the ensemble on 4,785 randomly sampled BP-OSD successes, and it broke only 2 of them. On the failure side no sampling was needed, since every BP-OSD failure was decoded with the ensemble. There, the ensemble corrects 75–93% of failures on n=144 at p ≥ .003 and 34–43% on n=72 at p ≥ .002, and a correct correction is in the candidate pool for 92–100% of failures.

We then recomputed the interval on the net improvement so that it accounts for all three sources of uncertainty: the baseline failure rate, the rescue rate and the damage rate, each with exact Clopper–Pearson bounds and a Bonferroni correction. At n=144 and p=.005, LER falls from 3.34% to 0.84%, a certified net gain of 1.43–3.05 percentage points at 95% confidence. The paper now reports the exact rescue fractions together with this certified result.

Q2. A fixed threshold chosen on separate validation data

The gate scores each sample by its predicted risk of failure, and in the paper it routes the top 10% of each batch to the ensemble. To remove the dependence on the batch, we fixed a single score threshold on separate n=288 calibration samples and applied it unchanged to the held-out codes at p=.10 (Appendix C, Table 3). Any sample scoring above the threshold goes to the ensemble; the rest keep the single PRIME decode.

CodeRouted to ensembleLER, fixed thresholdLER, ensemble on every sampleCalls per sample, fixed threshold vs. every sample
n=36015.7%0.90%0.90%7.01 vs. 34
n=7565.1%0.06%0.06%3.63 vs. 34

The two LER columns are identical: the threshold already routes the samples where the ensemble makes a difference, so also running it on the low-risk remainder adds nothing. The fixed threshold therefore gives exactly the full-ensemble accuracy at 5–9× fewer solver calls, on codes it was not calibrated on.

We thank the reviewer again for these suggestions, which made the circuit-level results more precise, and we are glad to provide anything further during the discussion.

Reviewer_vVoEOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The manuscript proposes PRIME, a hybrid decoder for quantum LDPC codes that uses a GNN to predict syndrome-conditioned, per-qubit error probabilities for an existing BP-OSD solver. These predictions replace the fixed channel priors used by the baseline while preserving the BP and OSD procedures. A single GNN trained on small bivariate bicycle (BB) codes is evaluated without retraining on larger BB codes and toric codes. The paper further introduces Jitter-PRIME, which generates multiple perturbed versions of the predicted priors and aggregates the resulting corrections by logical equivalence class under code-capacity noise. An optional logistic-regression confidence gate restricts this additional search to samples with the highest predicted failure risk. The experiments report lower logical error rates than BP-OSD-CS-6 and Astra in the evaluated BB settings. PRIME also achieves lower reported mean decoding times than BP-OSD-CS-6, using a BP iteration cap of 30 rather than 100. Jitter-PRIME further improves accuracy at the cost of additional solver calls, while selective routing reduces the average computational cost.

Critical strengths
  • PRIME and Jitter-PRIME shows good generalization capability. The manuscript's presentation is trained on ⟦36,8,4⟧,⟦72,12,6⟧,⟦144,12,12⟧\llbracket36,8,4\rrbracket,\llbracket72,12,6\rrbracket,\llbracket144,12,12\rrbracket, and generalize to ⟦288,12,18⟧,⟦360,12,24⟧,⟦756,16,34⟧\llbracket288,12,18\rrbracket,\llbracket360,12,24\rrbracket,\llbracket756,16,34\rrbracket while keeping good performance as presented in Figure 2.
Critical weaknesses
  • The latency measured in Figure 5 shows good speed gain, however, since PRIME is built on the basis of BP-OSD under 30 budget cap, an original BP-OSD solver with 30 budget cap could helps understand the performance gain of the method.
Questions for authors
  • PRIME provides syndrome-conditioned, per-qubit error probabilities to a classical decoding back end. Have the authors evaluated these learned priors with any compatible decoders other than BP-OSD? In particular, can the same trained GNN improve another decoder without retraining, or is decoder-specific fine-tuning required? Such experiments would help clarify the extent to which PRIME’s gains depend on its interaction with BP-OSD.

  • How do PRIME and the uniform-prior BP-OSD baseline compare under matched BP iteration caps and OSD configurations? In particular, could the authors report a uniform-prior baseline with a 30-iteration cap? It can help demonstrate the effectiveness of the design.

3 comments
AuthorsOpenReview ↗

Response to the Review

We thank the reviewer for the positive assessment and for helpful, precise questions. Following the reviewer's suggestions, we ran each requested comparison, and the results support the paper's claims. Section numbers refer to the revised PDF. Each table is paired on its own samples, so a setting that appears in two tables can differ slightly between them.

W1. Fig. 5 latency against BP-OSD at a 30-iteration cap

We timed uniform-prior BP-OSD at cap 30 with the Fig. 5 protocol (p=.10), and Fig. 5 now includes it:

CodeUniform BP-OSD, cap 30PRIME (incl. GNN)
n=2880.77 ms0.91 ms
n=3601.06 ms1.1 ms
n=7563.50 ms2.07 ms

With the learned prior, BP converges sooner, so the solver itself runs 1.9–2.8× faster than the cap-30 baseline. End to end, PRIME is close in speed on the smaller codes and 1.7× faster on n=756, with 3–8× lower LER on the same samples.

Q1. Do the learned priors improve another decoder without retraining?

Yes. We plugged the frozen GNN into Relay-BP (Müller et al., 2025), which has a different message-passing schedule and no OSD, with no fine-tuning (Appendix D.5, Figure 7). With its published settings on n=288:

pUniform BP-OSDRelay-BPRelay-BP + frozen PRIME priorPRIME
.081.33%0.61%0.35%0.22%
.094.96%2.49%1.49%0.93%
.1012.3%5.66%4.30%3.71%
.1122.7%15.6%10.2%7.62%

The prior alone lowers Relay-BP's LER by 24–42% at every rate, and a tuned Relay-BP profile improves the same way in all four settings, including the held-out n=360. An unmodified checkpoint thus transfers to a decoder it was never trained with.

Q2. Uniform-prior BP-OSD at matched BP caps and OSD configuration

We compared both priors at matched caps, with identical OSD-CS-6 and the same samples (Appendix E.5, Table 12). The learned prior has lower LER in every setting:

CodepCap 30: uniform / learnedCap 100: uniform / learned
n=288.070.32% / 0.10%0.42% / 0.10%
n=288.1010.8% / 3.5%10.4% / 3.2%
n=360.070.27% / 0%0.24% / 0%
n=360.108.7% / 2.2%8.0% / 2.1%
n=756.080.15% / 0%0.24% / 0%
n=756.104.6% / 0.34%4.5% / 0.44%

The gain therefore comes from the prior, not the cap. We also compared against a stronger, correlation-aware classical decoder that conditions one sector's prior on the other sector's correction (Section 5.2): at the same number of solver calls, PRIME's LER is 1.6–5× lower on n=288 / 360 / 756.

We thank the reviewer again for these suggestions, which helped demonstrate the effectiveness of the design, and we are happy to answer any further questions.

Thanks for the authors' responses, I think Weak accept is enough for this paper.

We thank the reviewer for their positive feedback and for taking the time to consider our responses. We appreciate their thoughtful evaluation of our work.

Reviewer_TTMfOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
1(−1)
Presentation
2
Soundness
1(−1)
Read review
Paper summary

PRIME uses a GNN to predict syndrome-conditioned priors for BP-OSD. Jitter-PRIME adds prior perturbations, logical-class pooling, and a gate that routes difficult cases to an ensemble. The experiments report lower logical error rates and transfer to larger BB codes and toric codes. The circuit-level variant uses native DEM priors.

Critical strengths
  1. The reported LER reductions are substantial. A predictor trained on small BB codes also transfers to larger codes without retraining.
  2. Table 1 shows that the gate approximately preserves ensemble accuracy on held-out BB codes while reducing average solver calls from 34 to 5.2.
Critical weaknesses
  1. I find the methodological novelty limited. PRIME learns per-qubit priors with a GNN and a bitwise loss, leaving BP-OSD unchanged. Neural refinement of BP inputs [1] and perturbed-prior ensembles [2] already exist. The contribution is mainly their combination with learned routing. Missing comparisons with alternatives such as Relay-BP [3] also make the practical advantage difficult to judge. The circuit-level results use native DEM priors, so they do not validate the central GNN component.

  2. The joint X/Z design is not well justified. Although the network sees both syndrome types, it outputs two marginals per qubit, followed by separate decoding and logical-class selection. This can use cross-syndrome information, but does not retain the joint distribution needed for a joint logical-class decision. Why this design is preferable to separate predictors remains unclear. Class pooling introduces another approximation: it sums likelihoods over discovered candidates, so its scores depend on which representatives the search finds. The paper acknowledges these approximations but does not establish how they affect its claimed treatment of correlations and degeneracy.

Questions for authors
  1. With the same ensemble and routing budget, how much does the learned prior improve over native priors? How does the LER–latency trade-off compare with Relay-BP [3] under matched settings?

  2. What do joint X/Z inputs and positional encoding contribute? For the components already ablated, the 5–8% changes in PRIME LER are smaller than the 12–17% difference between two full-model training runs reported in Appendix E.2. Which architectural benefits are reproducible across runs?

  3. How does the bitwise BCE plus entropy loss relate to logical decoding performance? Does the entropy term improve final LER, beyond making per-qubit predictions more confident?

  4. Does class pooling help in the main setting with learned priors and 16 perturbations? Appendix E.4 instead tests native priors with 64 perturbations. How sensitive is the class ranking to which representatives the search discovers?

  5. Are there circuit-level results using the GNN prior itself? Section 4.6 uses native DEM priors and a different candidate-selection rule, leaving the learned component untested in this setting.

3 comments
AuthorsOpenReview ↗

Response 1/3 to the Review

W1. Novelty relative to [1] and [2]

We thank the reviewer for raising this point, as it lets us state the contribution more clearly. The core contribution is a learned prior that improves existing decoders rather than replacing them: the same frozen GNN weights, with no retraining or fine-tuning, lower the error rate of every decoder we plugged them into.

  • Relay-BP [3], which has a different message-passing schedule and no OSD: the frozen prior lowers Relay-BP's LER by 24–42% at every noise rate we ran with its published settings (n=288: 0.61% → 0.35% at p=.08, 2.5% → 1.5% at p=.09), and halves it on the held-out n=360 with a tuned profile (1.56% → 0.78% at p=.09). The network was never trained with Relay-BP (Q1b).
  • A correlation-aware (Y-aware) BP-OSD decoder, the strongest classical baseline we tested: at a matched 42 calls, learned candidates lower its LER to 1.8% on n=288 and 1.0% on the held-out n=360, against 2.3% and 1.4% with classical candidates (Q1a).
  • BP-OSD itself: 3–4× lower LER than the uniform prior at matched BP cap and OSD configuration on n=288 and n=360.

This makes the prior a reusable decoder input, and it generalizes along the other axes as well:

  • Across code sizes: trained on n ≤ 144, the same weights give 21.8× and 15.7× lower LER than BP-OSD-CS-6 on n=1500 and n=3000 (p=.12).
  • Across code families: on toric codes, the same weights improve on MWPM by 2.3–10.7× and match a per-code neural decoder, without fine-tuning.

The gain does not come from search, which separates it from perturbed-prior ensembles [2]: two learned calls beat a 34-call uniform-prior ensemble (on n=288, 1.0% vs. 2.1% LER at p=.09 and 8.7% vs. 13.1% at p=.11). The learned gate is a further cost optimization, keeping full ensemble accuracy at 5.2 instead of 34 solver calls.

On circuit level, the reviewer is right that Section 5.3 uses native priors. We answer this in Q5, and the section is now titled "Circuit-Level Decoding with Native Priors".

W2. The joint X/Z design

Joint inputs vs. separate predictors. We retrained a control in which each sector sees only its own syndrome, at identical parameter count. Its LER is 2.8–5.9× higher on every one of 3 seeds × 3 noise rates. Cross-syndrome information is a main source of PRIME's accuracy, and separate predictors cannot use it.

Joint logical-class decision. Scoring each X/Z candidate pair jointly (a Y counts once) and pooling within joint logical classes, at no extra solver calls, lowers LER further on n=288:

pPer-sector selectionJoint selectionFloor (correct class not in pool)
.090.49%0.32%0.29%
.115.2%3.8%3.5%

Joint selection comes close to the floor set by the candidate pool (Appendix E.4).

AuthorsOpenReview ↗

Response 2/3 to the Review

Q1a. Learned vs. native prior at the same budget

The learned prior has lower LER than the uniform prior at every budget we tested (Appendix E.5, Tables 12 and 13):

  • Single decode (p=.10): 3.5% vs. 10.8% on n=288 and 2.2% vs. 8.7% on n=360.
  • K=16 ensemble, uniform jitter scale tuned on separate samples (n=288): 0.49% vs. 2.1% at p=.09 and 5.2% vs. 13.1% at p=.11.
  • Top-10% routing, each prior with its own gate: 0.63% vs. 4.3% on n=360.

To make the comparison stronger, we also added a correlation-aware classical baseline, Y-aware BP-OSD. Standard BP-OSD decodes X and Z independently with a flat 2p/3 prior, which ignores that a Y error flips both. Y-aware BP-OSD decodes one sector, then sets the other sector's prior to 1/2 on qubits where the first correction acts and to (p/3)/(1−2p/3) elsewhere, the exact depolarizing conditionals. Its strongest version runs both sector orders with a second conditioning pass and keeps the lighter X/Z pair (8 calls). At equal calls, PRIME's LER is 1.6–2.5× lower (4.1% vs. 6.4% on n=288 and 2.1% vs. 5.3% on n=360, p=.10). We then added 34 candidates to the 8-call Y-aware decoder from each source, at a matched 42 calls (p=.11, Table 5):

Added candidatesn=288n=360 (held out from all tuning)
None (Y-aware BP-OSD alone)3.6%2.2%
Uniform-prior2.8%1.9%
Classical (jittered Y-aware prior)2.3%1.4%
Learned (PRIME)1.8%1.0%

Learned candidates beat classical ones on both codes, with the comparison rule fixed in advance, and an independent replication on n=288 gave the same ordering. The learned prior therefore adds accuracy even on top of the strongest classical decoder we tested.

Q1b. Relay-BP [3]

Following the reviewer's suggestion, we ran Relay-BP as released, with its published settings on n=288 (Appendix D.5, Figure 7):

p=.08p=.09
Relay-BP0.61%2.5%
Relay-BP + frozen PRIME prior0.35%1.5%
PRIME0.22%0.93%

PRIME is more accurate than Relay-BP, and the frozen prior, with no fine-tuning, also lowers Relay-BP's own LER at both rates. A tuned Relay-BP profile shows the same improvement in all four settings, including the held-out n=360. On latency, Relay-BP is designed for parallel hardware and is faster in our CPU software; PRIME's advantage is accuracy, and the prior composes with Relay-BP to improve it as well.

AuthorsOpenReview ↗

Response 3/3 to the Review

Q2. Which architectural benefits reproduce across runs

The reviewer's reading was correct: the previous variant checkpoints did not apply the component switches, so the 5–8% differences reflected run-to-run variance. We retrained every variant with three seeds and verified the switches in each saved checkpoint (revised Table 1 and Table 8):

RemovedPRIME LER vs. full model, 3 seeds (p=.09 / .10 / .11)
Joint X/Z input5.7–5.9× / 3.9–4.4× / 2.8–3.0×, every seed
Positional encoding0.99–1.09 / 0.96–1.08 / 0.93–1.07
Entropy term0.98–1.13 / 0.97–1.12 / 0.95–1.05

The joint X/Z input is a large, fully reproducible benefit. The pipeline ablation (gate, ensemble, prior) is unchanged.

Q3. BCE, the entropy term, and logical LER

BP-OSD consumes a per-qubit error probability in each sector, and bitwise BCE is a proper scoring rule for exactly that quantity. Across the 12 retrained models, held-out BCE on the unseen n=288 code ranks LER with Spearman ρ = .72–.81 at the three rates (p ≤ .008), so the training objective tracks logical performance (Appendix E.2). The entropy term acts as a light regularizer; its effect on LER stays within run-to-run spread, and Section 6 now says so.

Q4. Class pooling with learned priors and K=16

We ran this at the paper's setting (learned priors, K=16) on fresh n=288 samples (Appendix E.4).

Does pooling help? Pooling only matters on samples where the candidate pool contains corrections from more than one logical class; otherwise every rule picks the same class. These samples are common at higher noise, and there pooling helps: at p=.11 (22% of samples) it lowers LER from 5.6% (lightest single candidate) to 5.2%. At p=.09, where only 4% of samples have competing classes, the two rules tie.

How sensitive is the ranking to the representatives found? We reran the search with a completely new draw of all 16 perturbations, so it discovers a different set of representatives. The selected class changes on only 0.4% (p=.09) and 3.6% (p=.11) of samples, and LER is the same under both draws: the few changed decisions fix as many samples as they break. The final accuracy therefore does not depend on which particular representatives the search happens to find.

Q5. Circuit-level results with the GNN prior

Following the reviewer's suggestion, we tested the GNN prior itself in the circuit-level memory experiments of Section 5.3 at p=.005, in four ways on n=144 (Appendix D.7, Table 7):

  1. A GNN trained on the detector error model, predicting a bounded correction to each fault mechanism's native prior: 3.02% vs. 3.00% LER (learned vs. native).
  2. That GNN as an extra candidate next to the native decode, against an extra native-jitter candidate at equal calls: 2.20% vs. 2.30%.
  3. Our shipped code-capacity checkpoint, unchanged, as a per-round prior on the fault mechanisms that act as single-qubit errors: 3.40% vs. 3.46%.
  4. The same checkpoint as features for the routing gate, against matched adjacent-fired-check features: recall at 10% routing changes by −1.3 points, CI [−3.7, +1.6].

In every case the learned prior performs on par with the native prior, and the shipped checkpoint does so without any circuit-level training. At circuit level the syndrome already locates most faults (simple adjacent-check counting reaches AUROC .95), which leaves little room for any prior, so the circuit-level gains in Section 5.3 come from the ensemble and gate. The section is titled accordingly.

We thank the reviewer again for questions that strengthened the paper, and we are glad to run further checks during the discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.