acceptodds
Under review as a conference paper at ICLR 2027

Two-Timescale Adaptive Decoding for Quantum Error Correction with Surface Codes

Abstract

A real-time Quantum Error Correction (QEC) decoder should adapt to complex and evolving hardware noise while remaining sufficiently lightweight to meet stringent latency requirements. To decouple adaptivity from latency while enabling label-free runtime adaptation, we propose , a slow-fast two-timescale decoding framework. SF-QEC employs two structurally aligned recurrent-transformer models inspired by AlphaQubit 2: a neural decoder operates on the fast path for streaming decoding, while an adaptation model operates asynchronously on the slow path to analyze longer histories of detection events. The slow model extracts history-derived context and transfers it to the corresponding recurrent states of the fast decoder, augmenting its temporal representation without requiring runtime labels or placing the slow model on the latency-critical path. On QEC data collected from the IBM Phoenix superconducting quantum processor, the full SF-QEC framework further improves already-superior decoding performance compared to baselines, demonstrating the potential of two-timescale neural decoding under evolving hardware conditions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
quantum error correction, neural decoder

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_HCjmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces SF-QEC, a QEC decoding framework designed to handle two requirements - adaptivity and latency. The proposed method decouples the two requirements to two paths, slow and fast, both inspired by AlphaQubit 2 recurrent-transformer architecture. A decoder operates on the fast path for streaming decoding, and on the slow path a slow model extracts longer history context, the history context is transferred and fused through state fusion. The framework is evaluated on experimental distance 3 surface code data only vs MWPM, Correlated Matching, Union-Find, BP-OSD / BP-LSD, and DGR alignment / Heuristic. The framework is also evaluated vs AQ2-RT. I find the main contribution of this paper to be the slow-to-fast augmentation mechanism, which I think can be used beyond this specific architecture.

Critical strengths
  • The new framework does improve decoding performance in most cases.
  • ⁠The separation into slow and fast paths is well motivated.
Critical weaknesses
  • The evaluation is very limited, the method is evaluated only on distance 3 surface code, it is hard to assess whether these improvements persist on a larger and more relevant scale.
Questions for authors
  • Could the authors provide evaluation on larger distance surface codes (for example 5, 7…) ?

  • The authors mentioned that SF-QEC has an additional joint training step, could the authors clarify if AQ2-RT baseline also does another training stage? If not, could they provide a comparison with additional trained AQ2-RT? Without it, it is hard to determine the source of improvement.

  • Could the authors clarify the reasoning for excluding batch 16 while including batch 8? It is unclear what the general criterion is to determine if a batch should be included or not. Could the authors also provide the results for batch 16?

2 comments

We sincerely thank the reviewer for their patience and insightful feedback.

Q1. Larger distance.

A1. We extended our experiments to distance-5 and distance-7 rotated surface codes with 10 syndrome rounds, using fast/slow widths of 64/64, a 500-shot history, and a 20-shot grouping and refresh interval. Synthetic data is generated with spatially varying sinusoidal noise. Starting from a Detector Error Model (DEM) with base circuit noise strength pp, we assign error mechanisms to four spatial regions according to their detector coordinates. Each region’s mechanism probabilities follow a sinusoidal multiplier with amplitude 0.6, giving an unnormalized range of 0.4–1.6 times their original probabilities. The results are provided below, showing lower observed LER in all four settings.

Code distanceBase noise strength ppAQ2-RTSF-QEC
50.0032.4942.150
50.0068.1998.026
70.0031.9861.955
70.00610.36410.278

Q2. Matching training budget.

A2. All AQ2-RT baselines reached early stopping before the 50-epoch budget cap, after five consecutive epochs without improvement in validation loss.

Rounds333366669999
Patch012301230123
Width 64272836334633463919423428
Width 128341826313237313419242026

We additionally trained each AQ2-RT baseline for the same number of epochs and optimizer updates to match corresponding SF-QEC joint-training budget without early stopping, and no improvement is observed, confirming that the fairness of our training budget.


Q3. Results for batch 16.

A3. Batch 16 shows severe degradation for both MWPM and the neural decoders, with LER approaching 50% (random guessing) at r=6 and r=9. The table reports LER (%) averaged across patches 0–3, using 8,000 test shots per batch and width-64 neural models.

RoundsTest batchesMWPMAQ2-RTSF-QEC
3Other batches8.5487.3757.096
3Batch 1634.08835.58835.309
6Other batches16.53414.49314.230
6Batch 1647.67847.50047.228
9Other batches23.07920.87120.706
9Batch 1649.00948.95649.066

The degradation is likely due to severe experimental condition change, and we hereby report the results separately.

Thank you for the response and the additional experiments. The rebuttal was helpful, but it did not significantly change my overall assessment. I find the evaluation relatively limited and think that the paper would benefit from a clearer presentation of the method. I will therefore maintain my original score.

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

This paper proposes SF-QEC, a two-timescale adaptive neural decoding framework for surface codes. The method combines a fast AQ2-RT decoder with a slow history encoder that extracts contextual information from previous detection events. The extracted information is incorporated into the fast decoder through recurrent-state fusion without requiring runtime labels. Experiments on IBM Phoenix demonstrate improved decoding performance over standalone AQ2-RT and conventional decoders.

Critical strengths
  • The slow-fast architecture is an interesting approach to incorporating long-term noise information without requiring online retraining or runtime labels.

  • The method demonstrates consistent decoding improvements over AQ2-RT on real quantum hardware data across multiple experimental settings.

Critical weaknesses
  • The real-time feasibility of the proposed framework is not fully demonstrated. The slow-path computation introduces substantial latency, while the evaluation assumes that updated slow states are immediately available to the fast decoder.

  • The improvement over standalone AQ2-RT is relatively modest, and it is unclear how much of the improvement comes specifically from the slow-path context rather than additional joint training.

Questions for authors
  • The slow path requires approximately 600–800 μs to process historical data, but the evaluation assumes that the updated context is immediately available for the next decoding block. How can this be achieved without delaying the fast decoder? Does the performance improvement remain when delayed slow states are used?

  • The improvement over AQ2-RT is relatively small, and joint training also updates the fast decoder. Could the authors compare against a fast-only model with the same additional training budget to isolate the contribution of the slow path?

1 comment

We sincerely thank the reviewer for their patience and insightful feedback.

Q1. Realtime feasibility

A1. We tested this delay using the original SF-QEC-64 checkpoints, without retraining, across all 12 IBM settings: four patches and 3, 6, or 9 syndrome rounds, using the paper’s same test shots. All SF-QEC columns use a 500-shot history and a 20-shot refresh interval.

RoundsPatchAQ2-RTSF-QEC: no delay20-shot delay40-shot delay100-shot delay
3011.17010.87710.93410.94810.978
313.6723.4313.4273.4393.450
326.6736.4506.4636.4946.478
337.9867.6277.6387.6447.677
6020.50520.19420.19520.22720.283
618.0207.7667.7707.7487.775
6214.80014.64214.66114.68314.656
6314.64814.31914.32714.30314.336
9028.72728.45528.48428.48328.516
9112.34812.12712.14512.11112.138
9221.60321.60321.60321.60321.603
9320.80820.64120.63020.65620.638

At every tested delay, SF-QEC retains a lower LER than AQ2-RT in 11 of 12 settings and ties in 1 setting. Although delayed context generally reduces the gain, even a 100-shot delay preserves an improvement wherever the original model improved over AQ2-RT.


Q2. Improvement justification.

A2. We agree that the benefit is not uniform. SF-QEC is intended to help when recent history reveals noise conditions that an individual shot does not reliably identify.

Using recent 500-shot histories, we measured increases in Z-check detector firing compared with their usual rates during training. We call the firing rate “elevated” when this increase, averaged over a batch, exceeds 90% of historical observations for the same patch (e.g., patch 0) and round settings (e.g., r=3). For each model width, we evaluate 3 round settings (3, 6, and 9 rounds), 4 patches, and 8 test batches, giving 3x4x8=963x4x8=96 batch evaluations. Among these, 12 batches have Z-check firing rate elevated. The table reports mean LER reduction over AQ2-RT in percentage points; larger values indicate greater benefit.

Z-check firing rateBatch evaluations per widthSF-QEC-64SF-QEC-128
Elevated120.5100.543
Other840.1970.177

This indicates large improvements can be obtained with SF-QEC when high Z-check firing rate occurs.


Q3. Matching training budget.

A3. All AQ2-RT baselines reached early stopping during experimental fine-tuning before the 50-epoch budget cap, after five consecutive epochs without improvement in validation loss.

Rounds333366669999
Patch012301230123
Width 64272836334633463919423428
Width 128341826313237313419242026

We additionally trained each AQ2-RT baseline for the same number of epochs and optimizer updates to match corresponding SF-QEC joint-training budget without early stopping, and no improvement is observed, confirming that the fairness of our training budget.

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

There are 2 constraints in quantum error correction problems. One is latency which is a tight constraint for runtime adaptation and the other one is labels which is only available in controlled experiments. To solve these 2 problems, the paper proposes SF-QEC, a two-timescale decoding framework consisting of a slow path and a fast path. A neural decoder operates on the fast path to process individual QEC cycles, while an adaptation model operates asynchronously on the slow path to analyze longer histories of detection events and provide contextual information to the fast decoder.

Critical strengths

The mechanism is deployed on Real-time AlphaQubit 2 architecture, the slow model captures history context through hierarchical history embedding, and transfers the context to the fast decoder through state fusion, without requiring runtime labels.

The framework evaluated on experimental surface-code data obtained on IBM Phoenix superconducting quantum computer, demonstrating further improved decoding accuracy compared to standalone Real-time AlphaQubit 2 under nonstationary noise.

Critical weaknesses

The hierarchical embedding compresses long histories into shot level and history level representations, this will reduce computational complexity. But it may lead to information loss. The adaptation performance maybe influenced by compression ratio and history window size.

Questions for authors

The framework contains several components, including hierarchical history embedding, fast and slow-path recurrent transformers, recurrent state fusion. Can you provide ablation studies to quantify each contribution of these components? It seems that the slow path has not online parameter updates, it’s unclear to me how this framework can generalize to noise distribution which are not presented during training. Figure 5 shows the slow state from slow path are final states of slow path and inputs of recurrent block, if they are running asynchronously, are the 2 paths updating at different frequency? How to coordinate slow path and fast path. What’s the conclusions of figure 8, 9, 10 of testing different Patch size and Bench Size? Can you explain the training strategy on line 286-291?

1 comment

We sincerely thank the reviewer for their patience and insightful feedback.

Q1. Ablation studies

A1. We evaluated three targeted replacements on IBM Phoenix distance-3 patch 0 at 3, 6, and 9 syndrome rounds, using fast/slow widths of 64/64, a 500-shot history, and a 20-shot grouping and refresh interval. The table reports LER (%) on the paper's eight test batches, including three ablations studying hierarchical embedding, recurrent-transformer processing, and recurrent-state fusion. Since all components are on the same path that constructs the slow path, instead of totally removing individual components, we replace them with simpler replacements.

(1) Hierarchical embedding is replaced with a two-layer MLP with GELU activation in between, hidden width 16. The replacement directly processes embeddings directly concatenated throughout the 20 shots without hierarchy.

(2) Recurrent-transformer processing is replaced with a spatial MLP. Specifically, we replace the first recurrent block and its two following spatial transformer layers, while retaining the second recurrent block.

(3) Recurrent-state fusion connections are replaced with a single feature-level fusion operation: the slow path delivers the final output features of its second recurrent block. These are concatenated with the corresponding fast-path per-stabilizer features and merged through a linear layer once per decoded shot, before the final spatial transformers and readout.

Modelr = 3r = 6r = 9
Full SF-QEC10.87720.19428.455
Replacing hierarchical embedding10.87820.20928.497
Replacing recurrent transformer10.95020.47328.839
Replacing recurrent state fusion10.98820.31328.613

Q2. Generalizing without online parameter updates

A2. We do not claim that SF-QEC generalizes to arbitrary noise distributions unseen from training. Instead, we hypothesize that a device's noise has learnable temporal structure that persists across many shots. Through offline training rather than online parameter update, SF-QEC allows the slow path to recognize the temporal structure and provide adaptive context for the fast path. The core obstacle towards online parameter update is: ground truth labels are generally unavailable during quantum computation. Although estimated labels may support online training, their reliability depends on noise condition assumptions.


Q3. State updates and asynchronous coordination

A3. The reviewer is correct that the two paths update at different frequencies. Below are details of state updates.

Within the slow path, shots 0-499 are embedded into 25 embedding blocks (20 shots per block), and each embedding block enters following recurrent blocks, updating the recurrent states 25 times. The “final state” means the final updated recurrent state for the 500 shots, which is used for state fusion with the fast path. After 20 shots, the slow path recurrent states are reset, and 500-shot window rolls forward, processing shots 20-519.

Within the fast path, every 6 rounds are embedded into an embedding block, and the fast recurrent states update once per round group and again when processing the final boundary. During the fast path update, the slow recurrent states remain unchanged, representing the same context of recent 500-shot history. After new slow recurrent states arrive, the cached slow states are replaced after every 20 shots.


Q4. Interpretation of Figures 8-10

A4. Figures 8–9 show similar average gains at both widths, suggesting that SF-QEC’s benefit persists even with a larger fast decoder. This indicates the slow path brings in new information beyond a larger fast model can carry. Figure 10 shows that these gains are uneven, with stronger improvements in batches 14–15. This indicates that the benefit is conditional rather than universal.


Q5. Training strategy

A5. SF-QEC is trained in three steps: (1) Train the fast decoder on simulated data, learning from abundant labeled examples to establish a backbone. (2) Fine-tune the fast decoder on experimental data, adapting to real-device noise that simulation may not capture. (3) Add the slow path and fusion layers, then train all three components together on the same experimental data as a united model.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.