acceptodds
Under review as a conference paper at ICLR 2027

Growing Dynamic Decoding Graph for High-Accuracy Quantum Error Correction

Abstract

Quantum error correction requires high-fidelity error priors to improve the decoding accuracy. Existing approaches mainly focus on estimating the probabilities or weights of decoding priors, while largely neglecting their graph topology, which limits the error correction performance even when the edge weights are well estimated. This motivates us to ask whether the decoding graph itself can be automatically constructed from syndrome samples and a given decoder. To this end, we introduce the PrunerNet that applies physics-guided sparse self-attention over the hyperedges in the detector error model. Each hyperedge maintains a learnable embedding to determine its retention probability, and attention is restricted to pairs of hyperedges that share stabilizer check components, capturing physically meaningful correlations in the noise model. By growing dynamic decoding graphs with reinforcement learning and under the direct supervision of the logical error rates, we achieve high-precision decoding priors without human-engineered calibration, enabling high-fidelity quantum error correction. Benchmarking on error correction circuits of the surface code demonstrates the effectiveness of learning dynamic decoding graph structures compared with leading methods that optimize prior probabilities alone. Across correlated-error and randomly reordered-circuit benchmarks and relative to fixed-topology calibration, our method averagely reduces logical error rates by and , respectively. We have also developed Auto-DEM, a modular compilation framework for detector error model, which enables joint optimization of the weights and topology of arbitrary decoding graphs, paving the way towards fault-tolerant quantum computing. All codes and evaluations are available at the https://anonymous.4open.science/r/Auto-DEM-ICLR-2027, and will be open-sourced upon publication.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum Error Correction, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_SzuHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a new method for optimizing the DEM in order to improve quantum error correction decoding. The Auto-DEM framework consists of three main components: a decoding graph compiler, PrunerNet and a reinforcement learning strategy. PrunerNet applies physics-guided sparse self-attention over the hyperedges in the DEM. First, the decoding graph is extended, then edges are selected using a sampled binary mask, and finally retained edges are weighted. PrunerNet is trained with reinforce under LER supervision. The proposed method is evaluated on the surface code with $ d = 3, 5, 7 $ over 3 syndrome extraction rounds. The two decoders used for the experiments are MWPM and Correlated Matching. Across both correlated-error and randomly reordered-circuit, relative to fixed topology calibration, the method averagely reduce LER.

Critical strengths
  • Optimizing the DEM with Reinforce under LER is a well motivated design choice.
  • The extension of the decoding graph followed by pruning and weight estimation is an interesting approach.
Critical weaknesses
  • The evaluation is limited to one code family and relative small code sizes. There is no evidence these improvements persist in more relevant settings, for example, surface code with $ d=9, 11....$
  • Only one noise level is tested. Without showing results over various noise levels, it is hard to determine whether the improvement is consistent.
Questions for authors
  • To better assess the generalization of this method, could the authors provide an additional evaluation on other code families (for example, BB codes)?
  • I have concerns regarding the scalability of the proposed method, could the authors provide some evaluation where the number of rounds matches the code distance for d=5 and d=7? Also, can the authors provide an evaluation on larger distance surface codes?
  • Since the model was trained with one physical error rate, could the authors provide some clarity regarding its generalization? For example, do the empirical results persist on other noise levels? or does the model need to be retrained for each physical noise level?
3 comments

We thank the reviewer for the constructive feedback. We conducted additional experiments on BB codes, longer surface-code circuits, and generalization across physical error rates. We address these concerns in three comments.

W1 & Q1. Generalization to other code families: BB codes

Following your suggestion, we evaluated Auto-DEM on the [[144,12,12]] BB code (Gross Code) with r=3r=3, using BP+OSD-0 and BP+OSD-CS4 to decode LER. The code construction and CNOT schedule follow https://www.nature.com/articles/s41586-024-07107-7, with ℓ=12,m=6\ell=12,m=6, A=x3+y+y2A=x^3+y+y^2, and B=y3+x+x2B=y^3+x+x^2.

Unlike our surface code experiments, we extended the DEM with higher-order hyperedges (Manhattan threshold 3, orders = [2, 3, 4]). This introduced 5,4595,459 candidate mechanisms beyond the original 14,61614,616 mechanisms (3960 basic, 10656 non-basic). Under SI1000 noise (p=0.003p=0.003), we used 200k training shots and 100k validation shots. Training consisted of 1,5001,500 iterations, each evaluating 8 candidate graphs on 5k shared shots. The reward was the mean LER across all 12 logical observables.

After training, we evaluated all methods on 10 independent batches of 100k shots. We compared the circuit-compiled Ground truth DEM, the original circuit graph structure + Parity (Nominal parity), the full extended graph + Parity (Full parity), and PrunerNet + Parity. We also tested the learned DEM on BP+OSD-CS4 to assess whether it had overfit to the BP+OSD-0. LER results are reported as mean(std):

DecoderDEMMean LER
BP+OSD-0Ground truth0.0170(3)
BP+OSD-0Nominal parity0.0168(2)
BP+OSD-0Full parity0.0168(2)
BP+OSD-0PrunerNet + Parity0.0103(2)
BP+OSD-CS4Ground truth0.0075(2)
BP+OSD-CS4Nominal parity0.0077(1)
BP+OSD-CS4Full parity0.0076(1)
BP+OSD-CS4PrunerNet + Parity0.0049(1)

Compared with the circuit-compiled Ground Truth DEM, Auto-DEM reduced mean LER by 39.56% with BP+OSD-0 and 34.22% with BP+OSD-CS4. All 12 logical observables showed improvements under both decoders. Interestingly, the improvement persists when switching from BP+OSD-0 to BP+OSD-CS4 without retraining. This suggests that the learned topology can benefit different decoding configurations.

We also measured graph complexity and decoding time. The following mean decoding times were measured on 1010 independent datasets of 100k100k shots on a single RTX 4090:

DEMEdges (Error mechanisms)BP+OSD-0 (s)BP+OSD-CS4 (s)
Ground truth14616105.3317.6
Nominal parity14616105.3313.3
Full parity20075153.0427.5
PrunerNet + Parity805957.5183.8

Compared with Ground Truth, Auto-DEM reduced decoding time by 45.36% and 42.14%, respectively. These results provide evidence that Auto-DEM can generalize beyond surface codes and matching decoders. In particular, topology optimization can improve both accuracy and efficiency for approximate decoders such as BP-OSD. We thank the reviewer for suggesting this valuable extension.

W1 & Q2. Scalability to larger code distances and longer circuits

We agree that evaluating larger code distances and more realistic measurement rounds is important. We conducted exploratory experiments up to d=9d=9 under the standard SI1000 noise model, where Auto-DEM continued to show improvements. However, training costs increase substantially with code distance due to decoding latency per epoch in the RL training pipeline, preventing a systematic evaluation at d=9,11,…d=9,11,\ldots within our current computational budget.

A major limitation is that LER-based rewards become increasingly difficult to estimate accurately as logical errors become rarer. Improving reward efficiency and training scalability remains an important direction for future research.

To investigate longer measurement sequences, we additionally conducted a cross-round transfer experiment under the correlated-noise setting. We trained a five-layer template graph at d=r=5d=r=5, consisting of initialization, repeated-measurement, and final-measurement sections. The repeated-measurement structure is shared across rounds.

Training used 1,500 epochs, with 16 sampled masks and 10k reward shots per epoch. After training, we fixed both the learned topology and edge weights, preserving the two boundary sections while expanding the repeated-measurement section to r=7,9,11,13r=7,9,11,13.

Importantly, no target-round data were used for retraining or weight recalibration. For each round count, all methods shared 10 fresh independent batches of 100k shots.

RoundsStandardPrunerNetOracleLER reduction
50.0228(3)0.0205(4)0.0198(3)9.85%
70.0298(7)0.0269(5)0.0255(6)9.85%
90.0372(8)0.0331(6)0.0311(8)11.09%
110.0442(5)0.0394(4)0.0369(3)10.88%
130.0511(6)0.0455(7)0.0425(6)10.92%

The model trained only at r=5r=5 retained its improvement when transferred to r=7–13r=7–13 without retraining or recalibration. This demonstrates that the learned template can be extended to longer memory circuits under stationary noise.

We will clarify these scalability limitations in the revised manuscript.

W2 & Q3. Generalization across physical error rates

Thank you for raising this important question. Our additional experiments show that retraining is not required for every physical error rate within the tested range.

For each noise setting, we evaluated one PrunerNet model trained at p=0.003p=0.003, with d=5,r=3d=5,r=3, and fixed both its graph structure and all edge weights. Without any retraining or recalibration, we evaluated it at five physical error rates, p∈0.0015,0.002,0.003,0.004,0.005p\in{0.0015,0.002,0.003,0.004,0.005}.

For each pp, we generated 10 independent batches of 100k fresh test shots, shared by all methods. The random-ordering schedule was unchanged, while the correlated-noise strength followed the original setting q=0.2pq=0.2p.

For comparison, Parity used an additional 200k shots at each pp to recalibrate the Standard DEM, whereas PrunerNet used no data from the new error rates. LERs are reported as mean(sample standard deviation) across the ten batches, with ddof=1; means have four decimal places, and parenthetical digits refer to the last decimal place.

Random ordering

ppParityPrunerNetOracle
0.00150.0023(2)0.0017(1)0.0012(1)
0.00200.0045(2)0.0035(2)0.0027(1)
0.00300.0113(2)0.0099(2)0.0080(4)
0.00400.0225(4)0.0207(6)0.0178(5)
0.00500.0383(7)0.0364(8)0.0326(6)

Correlated noise

ppParityPrunerNetOracle
0.00150.0032(2)0.0028(1)0.0026(2)
0.00200.0059(1)0.0054(1)0.0052(1)
0.00300.0151(4)0.0138(3)0.0138(5)
0.00400.0288(5)0.0272(4)0.0272(5)
0.00500.0474(6)0.0455(7)0.0456(8)

At all five error rates, the fixed PrunerNet DEMs outperform Parity recalibrated at each pp.

These results demonstrate that Auto-DEM can retain its decoding advantage across a range of physical error rates without retraining or weight recalibration, despite being trained only at p=0.003p=0.003.

We will include these results and clarify the scope of this generalization in the revised manuscript.

Thank you for the additional experiments and detailed responses. The results are encouraging, and I believe the method has potential. However, my main concern still remains, I think the method is still limited in terms of scalability, and the complexity of the architecture is not well justified. Therefore, I will maintain my original score.

Reviewer_nPinOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper trains a Transformer to choose which edges to keep in a quantum-error decoding graph, using the logical error rate as feedback. The authors report lower error rates than fixed-graph baselines in several simulated settings. My familiarity with quantum error correction is limited, so this review focuses on the learning method and the experimental comparisons.

The paper studies changing the graph itself, rather than only adjusting weights on an existing graph. For randomized-order circuits, the reported average logical error reductions are 14.1% with the Correlation estimator and 12.6% with Parity, relative to their fixed-graph baselines.

The random-edge comparison is useful: adding random edges hurts, while the learned choices improve the reported results. This supports the value of learning which edges to use.

From a machine-learning perspective, I am not yet convinced that the edge selector needs a Transformer. A simpler method that learns one selection probability per edge is not compared at the same reward-evaluation budget.

I found it hard to judge how practical the search is beyond the tested settings, which stop at code distance 7 and three rounds. The paper gives the training budget but not training time or final decoding time.

Critical strengths

The paper studies changing the graph itself, rather than only adjusting weights on an existing graph. For randomized-order circuits, the reported average logical error reductions are 14.1% with the Correlation estimator and 12.6% with Parity, relative to their fixed-graph baselines.

The random-edge comparison is useful: adding random edges hurts, while the learned choices improve the reported results. This supports the value of learning which edges to use.

Critical weaknesses

From a machine-learning perspective, I am not yet convinced that the edge selector needs a Transformer. A simpler method that learns one selection probability per edge is not compared at the same reward-evaluation budget.

I found it hard to judge how practical the search is beyond the tested settings, which stop at code distance 7 and three rounds. The paper gives the training budget but not training time or final decoding time.

Questions for authors

What advantage does the Transformer have over learning one selection probability per edge?

How long do training and final decoding take at code distance 7?

1 comment

W1 & Q1. Independent logits and the role of the Transformer

Thank you for suggesting a comparison with independently learned edge-retention probabilities under the same reward-evaluation budget. We have added this baseline and conducted further architectural ablations.

Independent logits baseline. Following your suggestion, we replaced PrunerNet with a trainable vector of independent edge logits, removing the embeddings, Transformer, and MLP. Each logit parameterizes the retention probability of one candidate hyperedge. We sampled edge masks from these probabilities, estimated edge weights using Parity, and updated the logits using the same LER-based REINFORCE objective as PrunerNet.

Both methods used the same candidate graph, training and validation datasets, Correlated-Matching decoder, initial retention probability of 0.9, and reward-evaluation budget. Experiments were conducted at d=5,r=3d=5,r=3, and the final LERs were evaluated on 10 independent datasets of 100k shots each, shared across methods.

MethodRandom orderingCorrelated noise
Independent logits + Parity0.0123(4)0.0148(4)
PrunerNet + Parity0.0099(3)0.0139(3)

PrunerNet achieved approximately 19.5% and 6.4% lower LER than Independent logits in the two settings, respectively. These results suggest that learning edge-retention probabilities through a shared neural network provides an advantage over optimizing independent parameters for individual edges under the same reward budget.

Architectural ablations. We further evaluated individual network components under the standard SI1000 noise model. All models shared 10 independent batches of 100k shots.

ConfigurationLER
Full PrunerNet0.00708(23)
Full attention0.00714(20)
Random attention0.00710(20)
MLP without attention0.00723(25)
Independent logits0.01823(47)

The ablations indicate that the intersection-based attention mask and attention layers provide relatively modest additional improvements, while the MLP-only model already achieves performance close to the full PrunerNet. Thus, although the Transformer improves decoding accuracy, the main benefit appears to arise from shared neural-network parameterization rather than the Transformer architecture itself. For applications where training cost is a primary concern, a simpler MLP-based architecture may offer a practical alternative.

W2 & Q2. Training and decoding time

Thank you for highlighting the missing runtime measurements. We provide the available training-time estimates and additional decoding-time measurements below.

Training time. At d=7,r=3d=7,r=3, the average training time was approximately 10–12 hours on an Intel Xeon Platinum 8358P CPU server. Precise timing logs were not retained. At d=5d=5, the average training times were 1.26 hours for random ordering and 1.92 hours for correlated noise.

The current implementation has not been optimized for training speed. Its main bottleneck is repeatedly invoking the decoder to evaluate LER-based rewards before each gradient update. Since graph optimization is performed offline, this training overhead does not directly affect the latency of the final decoder.

Final decoding time. Compared with the extended graphs, PrunerNet reduces mean decoding time by 38.96% and 21.95% for random ordering and correlated noise, respectively. Detailed graph sizes, mean decoding times, and measurement conditions are provided in our response to Reviewer DT7r, W1.

Finally, inspired by the BP-OSD experiments on BB codes discussed in our response to Reviewer SzuH, we used a BP-OSD decoder supporting parallel GPU decoding for surface-code training, reducing the training time at d=5d=5 to approximately 0.7 hours while producing DEMs with LER performance comparable to those trained using PyMatching.

Reviewer_5yrXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for learning decoding graph topologies for quantum error correction using reinforcement learning which minimises logical error rates. They introduce Auto-DEM, enabling joint optimisation of edge weights and decoding graph topology under a policy to minimise the logical error rate for a given code and decoder. This is facilitated through extending the DEM and partitioning it into a basic graph and learnable edges. They introduce the transformer model PrunerNet, consisting of L transformer layers and H attention heads, which learns the relative importance of learnable hyperedges and keeps/removes them. The paper evaluates Auto-DEM on d=3,5,7 (r=3 extraction rounds) surface code memory circuits under SI1000 circuit-level noise at physical noise p=0.003. They evaluate this setup with both correlated errors mediated by a nearest-neighbour exchange channel, and with random syndrome extraction schedules, two forms of decoding-prior mismatch which cannot be captured by the standard DEM. The learned topology reduces logical error rates by 3.8% and 13.4% respectively across these two settings, on average, relative to the tested fixed topology baselines. Ablations are performed to examine the effect of model size, extension order, entropy coefficient and reward sample size.

Critical strengths

The core idea of learning DEM topology with LER minimisation as the target is a novel direction and of interest to the community.

The method is modular so that a variety of estimator-decoder pairs may be used.

The LER reduction in the evaluated settings is larger than that of the tested fixed topology methods on average, with gains particularly noticeable in the random-ordered stabiliser benchmark.

Critical weaknesses

The benchmarks seem somewhat arbitrary and of limited applicability, in particular regarding random-ordered stabiliser measurements, it is unclear what practical advantage there is to the method here rather than compiling the standard circuit and calibrating that. Further, correlated noise is only examined from an injected channel which is local, nearest neighbour and same-round, other errors such as leakage and non-local mechanisms which are used as motivation are not tested. Furthermore, topology selection is a special case of weight learning on G_ext as setting the weight to 0 removes the edge. Natural Comparisons all apply to the extended graph: syndrome based estimation (noting Figure 6), Maximum-likelihood estimation, and LER-supervised weight estimation.

The method is tested only on small circuits for d=3,5,7 surface codes with 3 syndrome extraction rounds. Scaling is a concern and remains largely unexplored. This also raises the concern regarding the method's applicability to high-rate qLDPC codes, which remains unexamined despite claims of being applicable in principle to arbitrary codes. The extended graph adds edges of order at most 2, so augmenting hyperedges of higher order remains open. Hypergraphs are stressed throughout the paper but only matching decoders are examined, adding higher-order mechanisms and examining decoders such as BP-OSD or Tesseract would address the hypergraph claims in more detail.

While the concept of DEM topology optimisation targeting LER minimisation is a novel and interesting direction, the evidence of the efficacy of the model remains limited as it has only been tested on very small circuits (d=3,5,7, r=3) circuits for only the surface code. Demonstrated performance at larger scales and more general codes would strengthen the paper. Furthermore the tested settings seem somewhat arbitrary and do not adequately address realistic scenarios as is motivated in the paper, requiring further justification. In summary, I believe the work has substantial potential to be of interest to the community, however at present it has not demonstrated utility in the required settings of interest.

Questions for authors

Topology selection is a special case of weight learning on G_ext as setting the error probability to 0 removes the edge. Natural Comparisons all apply to the extended graph: syndrome based estimation (while acknowledging Figure 6), Maximum-likelihood estimation, and LER-supervised weight estimation.

Statistics are unclear: the reported non-oracle standard deviations of Table 1 are approximately 3 times the binomial shot noise for 100k shots, further clarification of the 10 repetitions and how these figures are arrived at is required. How are these figures arrived at?

PrunerNet gives each edge a free embedding with no explicit input features, this puts into question if the network matters. Geometric features that may allow transfer (e.g. train at r=3, use at r=d (as is typical)), remain unexplored. Does a topology learned at r=3 transfer to longer memories such as a typical r=d setting? Is the network necessary as there are no explicit input features, how does the pruner compare with independent logits?

How does training time scale with increased code and circuit sizes?

4 comments

We thank the reviewer for the detailed and constructive feedback. We conducted additional experiments on LER-supervised weight optimization, qLDPC codes, cross-round generalization, and training scalability. We respond to each concern below in three comments.

W1. Practical relevance of the experimental settings

We agree that recompiling the actual stabilizer circuit with Stim is preferable when the circuit and noise mechanisms are accurately known. Our random-ordering experiment instead examines a extreme mismatch scenario, where the decoder’s Standard DEM differs from the executed circuit and must be adapted using experimental data.

Similarly, the correlated-noise experiment evaluates whether Auto-DEM can identify missing error mechanisms beyond Standard DEM. We acknowledge that our injected nearest-neighbor noise does not fully capture realistic leakage or nonlocal correlations. Such mechanisms may require more comprehensive noise simulator beyond Stim and remain important directions for future evaluation.

W2 & Q1. Topology selection and weight optimization

We agree that topology selection can be viewed as a special case of weight optimization on the extended graph, where setting an edge probability to zero effectively removes that edge.

We therefore added three baselines: (1) Gaussian REINFORCE for LER-supervised weight optimization on both the Standard and Extended DEMs; (2) Independent logits for direct edge-selection optimization; and (3) Parity estimation on the Extended DEM with zero clipping.

All experiments used d=5,r=3,p=0.003d=5,r=3,p=0.003, with independent training and validation sets of 200k shots each. Final LERs were evaluated on 10 independent datasets of 100k shots.

MethodRandom orderingCorrelated noise
Standard Graph + Parity0.0112(4)0.0149(4)
Extended Graph + Parity (zero clipping)0.0597(10)0.0181(4)
REINFORCE + Standard Graph0.0101(4)0.0139(4)
REINFORCE + Extended Graph0.0097(4)0.0143(3)
Independent logits + Parity + Extended Graph0.0123(4)0.0148(4)
PrunerNet + Parity0.0099(3)0.0139(3)

These results show that continuous weight optimization on the extended graph can achieve comparable decoding accuracy to explicit topology selection. However, retaining all candidate edges may introduce unnecessary decoding overhead, even when their optimized probabilities are small.

Interestingly, we also found that Auto-DEM produces more compact graphs and enables faster decoding than LER-supervised weight optimization on the full Extended DEM. Detailed graph sizes, mean decoding times, and measurement conditions are provided in our response to Reviewer DT7r, W1.

Thus, while topology selection is mathematically related to weight optimization, explicit topology learning offers a useful accuracy–complexity trade-off compared with retaining and optimizing the full extended graph.

For maximum-likelihood estimation such as dMLE (Cao et al., 2026), we have evaluated it at d=3,5d=3,5. On random ordering, dMLE achieved LERs of 0.023115/0.010700, compared with 0.018945/0.010405 for Auto-DEM. However, dMLE was evaluated only on the Standard DEM because extending its tensor-contraction formulation to the expanded graph is computationally challenging. Training at d=7d=7 has already exceeded our GPU memory capacity due to the computational cost of tensor contraction.

W3. Generalization to qLDPC codes and higher-order hyperedges

To examine applicability beyond small surface codes, we conducted new experiments on the [[144,12,12]] BB code with r=3r=3, using the BP-OSD decoder and higher-order error hyperedges.

The original restriction to order-2 edges was imposed by the matching-based decoders used in our surface-code experiments, rather than by Auto-DEM itself. In the BB-code experiment, we extended the candidate graph to include higher-order hyperedges (Manhattan threshold 3, orders = [2, 3, 4]) and optimized their selection using PrunerNet.

Under the standard SI1000 noise model, Auto-DEM reduced the mean LER by 35% and decoding time by 40% compared with the Ground Truth DEM. The decoding-time reduction primarily results from the sparsity of the learned DEM.

Detailed results are provided in our response to Reviewer SzuH, W1 & Q1(BB codes).

These findings provide preliminary evidence that Auto-DEM can extend beyond surface codes and matching decoders. In particular, the results suggest that topology optimization can improve the performance of approximate decoders such as BP-OSD while reducing graph complexity.

We acknowledge that broader evaluations across qLDPC code families, code sizes, and higher-order error mechanisms remain necessary.

Q2. Standard deviations and repeated evaluations in Table 1

We thank the reviewer for identifying this inconsistency. After rechecking the original data, we found that different statistical protocols were inadvertently combined in Table 1. Specifically, the non-oracle results were calculated from 20 batches of 10k shots each, whereas the Oracle results used 10 independent batches of 100k shots. This difference explains the unexpectedly large standard deviations. We apologize for this mistake.

We have reconducted the evaluations using a consistent protocol:

ddDecoderStandardCorrelationParityPruner+Corr.Pruner+ParityOracle
3PyMatching0.0241(3)0.0240(3)0.0241(4)0.0188(2)0.0192(3)0.0194(3)
3Corr. Matching0.0236(4)0.0234(5)0.0240(6)0.0190(4)0.0192(4)0.0179(3)
5PyMatching0.0127(4)0.0119(4)0.0117(4)0.0106(4)0.0110(3)0.0098(3)
5Corr. Matching0.0124(3)0.0111(3)0.0112(4)0.0100(3)0.0103(2)0.0081(3)
7PyMatching0.0096(3)0.0088(3)0.0087(3)0.0079(3)0.0074(3)0.0067(2)
7Corr. Matching0.0079(2)0.0071(2)0.0071(2)0.0066(2)0.0067(2)0.0050(2)

We will replace Table 1 with these consistently evaluated results and correct the statistical description in our revision.

Q3. Independent logits and the necessity of PrunerNet

Following the suggestion, we implemented an Independent logits baseline by replacing the embeddings, Transformer, and MLP with a trainable retention logit for each candidate edge. Mask sampling, Parity estimation, and REINFORCE training remained unchanged.

At d=5d=5, Independent logits obtained LERs of 0.0123(4) and 0.0148(4) on random ordering and correlated noise, compared with 0.0099(3) and 0.0139(3) for PrunerNet. We further examined network components under the standard SI1000 setting. All models shared 10 independent batches of 100k shots.

ConfigurationLER
Full PrunerNet0.00708(23)
Full attention0.00714(20)
Random attention0.00710(20)
MLP without attention0.00723(25)
Independent logits0.01823(47)

These results suggest that shared neural parameterization contributes more than the specific attention architecture. Independent logits perform substantially worse, whereas removing attention causes only modest degradation. We therefore do not claim that the Transformer itself is indispensable.

Q3. Transfer to longer QEC circuits

To investigate cross-round generalization, we trained a five-layer template graph on correlated-noise data at d=r=5 and expanded the learned graph to longer circuits using the repeated block in Stim. The template contains initialization, repeated-measurement, and final-measurement sections. After training, both topology and edge weights were fixed, and only the repeated block was extended.

Detailed results from 10 independent batches of 100k shots per round are provided in our response to Reviewer SzuH, W1 & Q2.

The topology trained only at r=5 retains its advantage when transferred to r=7–13, without additional training or weight estimation.

This experiment demonstrates cross-round transfer ability due to the time-translation invariance of the memory experiment. However, training at r=3 is not suitable for template graph expansion, because a complete template graph requires at least five layers. We will add a new section in the Appendix to describe the graph extension method in our revision.

Q4. Training-time scalability

In our large-scale tests at d=5, training took an average of approximately 1.26 hours for the random-ordering task and 1.92 hours for the correlated-noise task. The average training time at d=7 and r=3 was approximately 10-12h, while an early exploratory training run at d=9 took approximately 36h. The current code has not been optimized for training speed. The main bottleneck is running the decoder to obtain a batch of LER values before each gradient update. Most mainstream surface-code decoding algorithms currently run primarily on CPUs. Deploying the entire algorithm on an FPGA or GPU platform with dedicated optimizations should provide substantial acceleration and greatly reduce the overall training time. Inspired by the BP-OSD experiments on BB codes, we used a BP-OSD decoder supporting parallel GPU decoding for surface-code training, reducing the training time at d=5 to approximately 0.7 hours while producing DEMs with LER performance comparable to those trained using PyMatching. This also shows the substantial potential to accelerate our training pipeline.

Reviewer_5yrXOpenReview ↗

Response to comments from authors.

I would like to thank the authors for their response and substantial additional experiments. I have read through them and the referenced responses to other reviews.

Table 1 is now consistent and the cross-round transfer result partly addresses my question, although models trained at r=3 cannot be transferred in this way. The training time details are appreciated.

Updated Assessment:

Topology vs weights: At d=5, LER-supervised weight optimisation is comparable to PrunerNet’s accuracy on the surface code, even without the graph extension on correlated noise. The gains may come from LER supervision, with topology selection mainly giving more compact DEMs than the extended graph, which I agree reduces overhead, however the net benefit once training cost is included may require further investigation.

Architecture: Full, random and no attention seem to have comparable performance, while independent logits perform substantially worse. The benefit appears to come from the shared parametrisation rather than physics-guided attention.

Realistic Evaluation: I agree with the authors that leakage is beyond Stim, however gate crosstalk and longer-range correlated measurement errors can be expressed as Pauli channels after twirling. I also note the Willow experiment, which shows no consistent advantage over Google’s RL prior across rounds, so realistic-noise performance remains an open question.

BB Code:For me this is the most compelling result of the work: 34-40% lower LER and about 42-45% faster decoding. As there is no prior mismatch and the learned DEM beats the ground-truth while removing genuine mechanisms, I read this as adapting the graph to suit BP-OSD rather than identifying missing noise mechanisms. Framing the work as LER-driven DEM sparsification for approximate decoding may better reflect its strengths.

Further Clarification Request (no new experiments required): Both PrunerNet and the independent-logit baseline start from the same initial retention probability (0.9) on the graph with the same reward budget. Were the optimiser settings tuned separately for the independent logits baseline, or shared from PrunerNet’s configuration?

Overall the additional experiments have improved the rigour of the paper and the BB-code result points to a promising contribution.

Thank you for this clarification request.

For the standard-SI1000 architecture ablation, the independent-logit baseline inherited PrunerNet’s optimizer settings and was not tuned separately: AdamW with an initial learning rate of (0.001), weight decay of (10−510^{-5}), and the same learning-rate and entropy schedules. Both methods used an initial retention probability of (0.9), REINFORCE updates, and the same reward-evaluation budget. Our original comparison therefore measured performance under shared optimization settings, rather than the best achievable performance of each parameterization.

Therefore, we further conducted a series of independent-logit experiments with the same updated budget on the SI1000 model in ablation study. Increasing the independent-logit learning rate to 0.01, 0.03 yield LERs of 0.01126(36), 0.00816(29), respectively, compared with 0.00708(23) for PrunerNet. No further decrease in LER was observed at higher learning rates.

Therefore, it cannot be overly assertive to claim that the neural network with shared parameters is necessary for this problem. It can only be said that the neural network with shared parameters can further reduce the LER by approximately 15% compared to the independent-logit method. We agree that the stronger supported contribution is LER-driven optimization of compact DEMs for approximate decoding, particularly in the BB-code experiment, rather than uniquely identifying the underlying physical noise mechanisms.

We will modify the corresponding conclusions of the independent-logit experiments in our revision paper.

Reviewer_DT7rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Auto-DEM, which learns both the topology and weights of the detector error model (DEM) used by QEC decoders. It extends the Stim-compiled DEM with nearby candidate hyperedges, uses a Transformer-based pruner (PrunerNet) to sample which candidates to keep, estimates edge weights with existing syndrome-statistics methods, and trains the pruner via REINFORCE using the decoded logical error rate (LER) as the reward. Experiments on simulated surface codes (d = 3, 5, 7) with injected correlated errors or randomized stabilizer order show lower LER than the standard DEM and weight-only recalibration.

Critical strengths
  1. The paper tackles an important and practically relevant problem in QEC decoding, presents it clearly, and offers a well-founded method.

  2. The idea of learning the structure of the decoding graph, instead of only tuning edge weights, is new and makes sense. The examples in Figures 1 and 2 (crosstalk and reordered stabilizer circuits) clearly show why a fixed graph might not be enough.

  3. The framework is flexible. It treats weight estimation and decoding as black boxes, so it can work with different estimators and decoders, and keeping a basic graph that is never pruned ensures decoding always works.

Critical weaknesses
  1. The paper only compares against Correlation and Parity. Sivak et al. isn't open-sourced, but the authors could try to reproduce it, and it's the most important comparison to make. Both methods train with LER as the signal; the only difference is that one changes the graph topology and the other changes the edge weights. Without this comparison, it's hard to tell whether the improvement comes from learning the topology or simply from using LER supervision. The authors also mention the MLE method of Cao et al. (2026) in the related work. I'd suggest adding it to the comparison as well.

  2. The paper's motivation is that real hardware has many errors the standard DEM doesn't capture, such as crosstalk and leakage. In the experiments, however, the correlated noise is added by the authors using a simplified model, which may be quite different from what actually happens on hardware. As far as I know, it is now possible to apply for access to real quantum hardware (e.g., IBM). Adding real-device experiments would make this paper much stronger.

Questions for authors
  1. I'm curious why PrunerNet only gives each edge a learnable embedding, without adding any information about the edge, like its coordinates, order, or estimated weight. Without that information, what does the Transformer add compared to simply learning a keep probability for each edge directly? Did the authors try that simpler approach?

  2. The ML components used here (REINFORCE with a baseline, entropy regularization, and a masked Transformer) are all fairly standard, so the paper's main contribution is on the QEC side rather than the ML side. I think the paper should explain more clearly. Like why is learning-based topology selection necessary? Could simpler search methods do just as well? What specific challenge in QEC makes this ML approach a good fit?

2 comments

We thank the reviewer for the constructive feedback. We conducted additional experiments on LER-supervised weight optimization, independent logits, local search, and Google’s Willow hardware data.

W1. Comparison with Google's LER-supervised weight optimization method and dMLE

We agree that the benefits of topology learning must be distinguished from those of LER supervision. Our Gaussian REINFORCE baseline is a simplified adaptation of the LER-supervised weight optimization approach of Sivak et al. (2024), omitting the multi-sensor optimization due to its implementation complexity.

All methods were evaluated at d=5,r=3,p=0.003d=5,r=3,p=0.003, using independent training and validation sets of 200k200k shots each. Final LERs were evaluated on 1010 independent datasets of 100k100k shots.

MethodRandom orderingCorrelated noise
Standard graph + Parity0.0112(4)0.0149(4)
Extended graph + Parity (zero clipping)0.0597(10)0.0181(4)
Gaussian REINFORCE + Standard graph0.0101(4)0.0139(4)
Gaussian REINFORCE + extended graph0.0097(4)0.0143(3)
Independent logits + Parity0.0123(4)0.0148(4)
PrunerNet + Parity0.0099(3)0.0139(3)

These results show that LER supervision itself contributes to the improvement, as REINFORCE achieves competitive accuracy even without topology learning. Interestingly, we also find that Auto-DEM achieves competitive LER with more compact graphs and faster decoding than weight optimization on the extended graph. We compared PrunerNet with Gaussian REINFORCE on both extended and Standard DEMs, reporting the number of DEM error mechanisms and mean decoding time.

Times are arithmetic means over 10 shared batches of 100k shots per task, using Correlated-Matching on an Intel Xeon Platinum 8358P with frozen DEMs.

Random ordering

MethodDEM error mechanismsMean time (s/100k shots)
PrunerNet + Parity7060.5391
Gaussian REINFORCE + extended DEM graph14550.8832
Gaussian REINFORCE + Standard DEM graph8610.6183

Correlated noise

MethodDEM error mechanismsMean time (s/100k shots)
PrunerNet + Parity8840.6039
Gaussian REINFORCE + extended DEM graph11130.7737
Gaussian REINFORCE + Standard DEM graph8610.6488

The optimized graphs reduce the mean decoding time by 38.96% and 21.95%, respectively, compared with the extended graphs. These results demonstrate that the learned graph structures can improve not only decoding accuracy but also decoding efficiency by reducing graph complexity.

Regarding dMLE (Cao et al., 2026), we evaluated it at d=3,5d=3,5. On random ordering, dMLE achieved LERs of 0.023115/0.010700, compared with 0.018945/0.010405 for Auto-DEM. dMLE was evaluated only on the Standard DEM because extending its tensor-contraction formulation to the expanded graph is computationally challenging. Due to the exponential complexity associated with tensor contraction, training at d=7d=7 has already exceeded our GPU (RTX 4090) memory.

W2. Evaluation on real hardware

We evaluated Auto-DEM on Google’s public Willow Z-memory data (d5_at_q6_9 for simplicity). Training used mixed rounds r∈10,30,50,70,90r\in{10,30,50,70,90} with a five-layer template graph whose repeated-measurement structure was shared across rounds. During training, the network maintained a five-layer template graph divided into initialization, repeated measurement, and final measurement sections. At each evaluation, the template was expanded to the specified rr. After training, the learned template graph is fixed and can be extended to any r≥5r \geq 5. To allow higher decoding fidelity, for each evaluated rr, edge weights were re-estimated using its training shots with the Parity.

The following LERs were evaluated on the held-out 10k shots using Correlated Matching. Google’s RL prior and SI1000 DEMs were provided with the dataset.

RoundsPrunerNetGoogle RL priorSI1000
100.04620.04070.0417
300.14140.15750.1559
500.19380.18510.1919
700.25490.23160.2518
900.32900.33350.3484

Auto-DEM outperforms both baselines at 30 and 90 rounds but shows no consistent advantage across all round counts. These results demonstrate applicability to hardware data while highlighting the need for improved training in future research.

The comparison with Google’s RL prior also requires caution. Their prior in Willow was optimized using a separate 13-cycle calibration dataset collected before the experiment (see SI of Willow's paper). We could not identify this dataset in the public Zenodo and therefore cannot determine its exact size or assess the training conditions.

Q1. Independent logits and the role of Transformer

We implemented the suggested Independent logits baseline, replacing our PrunerNet with a trainable retention logit for each candidate edge while the keeping mask sampling, Parity estimation, and REINFORCE training pipeline unchanged.

Independent logits for d=5,r=3d = 5, r=3 surface code obtained LERs of 0.0123(4) and 0.0148(4) on random ordering and correlated noise, respectively, compared with 0.0099(3) and 0.0139(3) for PrunerNet.

We also evaluated the contribution of different network components under the standard SI1000 setting. All models shared 10 independent batches of 100k shots; results are mean(sample SD, ddof=1), with SD digits aligned to the last decimal place.

ConfigurationLER
Full PrunerNet0.00708(23)
Full attention0.00714(20)
Random attention0.00710(20)
MLP without attention0.00723(25)
Independent logits0.01823(47)

These results suggest that shared neural parameterization contributes more than the specific Transformer architecture or attention mask. In particular, removing attention causes only a modest degradation, whereas independent logits perform substantially worse in this setting.

These results show that the complex adjacency mask and nonlinear attention layers make a relatively limited contribution to the final LER improvement. However, a neural network with shared parameters is necessary for the current graph-structure learning task. When training time and complexity are concerns, omitting the more computationally demanding Transformer is an option for future work.

Q2. Why learning-based topology selection? Can simple search suffice?

The central challenge is that decoding performance depends on the joint configuration of error mechanisms, rather than on the accuracy of individual edge probabilities alone. Missing error mechanisms cannot be explicitly represented by reweighting the original decoding graph, while indiscriminately adding candidate edges can degrade decoding, as demonstrated by the Extended DEM + Parity experiment in our original paper.

Moreover, the discrete decoding process prevents direct backpropagation through LER, and finite-shot LER estimates are noisy. REINFORCE provides a black-box optimization approach that can account for interactions among candidate edges without differentiating through the decoder.

To evaluate whether simpler optimization could suffice, we have also implemented a simple local-search baseline. At each step, it evaluates the current graph and 1515 randomly modified candidates, re-estimates their Parity weights, and retains the graph with the lowest training LER. The search uses the same number of reward evaluations and shots per evaluation as PrunerNet.

MethodRandom orderingCorrelated noise
Local search0.0111470.014905
PrunerNet0.0096200.013760

Under the same evaluation budget, PrunerNet outperforms this simple local-search baseline. Although RL with neural network may not be the only possible approach for DEM topology optimization, we believe these results support its effectiveness for navigating the search of candidate graph structures with high complexity.

Two QEC-specific challenges motivate our learning-based approach. First, physical error mechanisms in quantum hardware are complex and not fully characterized, making it difficult to construct accurate DEMs using predefined noise models. Data-driven learning can help capture error mechanisms that are difficult to model explicitly. Second, due to quantum error degeneracy and decoding heuristics, a more accurate DEM does not necessarily lead to better logical decoding. Finding a suitable graph topology for a particular decoder is therefore nontrivial. Our approach addresses these challenges by directly optimizing graph topology through LER feedback, without differentiating through the decoder.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.