quant-phOct 30, 2026
SaveDistilling Knowledge from Search: Decoder Model Factory for Quantum LDPC Codes
Organizations: EdenCode Inc., San Jose, CA 95134, USA · Department of Physics, University of California, San Diego, CA 92093, USA
Abstract
We introduce GraphMP, a learning-based message-passing decoder for quantum low-density parity-check codes that transfers without fine-tuning across code families, code sizes, physical error rates, and noise models. GraphMP uses local message passing to keep computation scalable as the decoding graph grows. Guided by its predictions, beam search progressively refines candidate error patterns, trading additional decoding time for lower logical error rates. We develop the decoder model factory as a paradigm for training successive models on experience generated by earlier models during decoding. Model-guided search supplies verified trajectories for knowledge distillation, folding computation spent on decoding into reusable training data. Repeated collection and distillation improve the decoder across generations and adapt its strategy as the noise changes. These results motivate a coupled quantum--classical architecture with two interacting cycles: rapid decoding protects quantum information, while slower learning renews the models from operating experience. The long-term vision is a scalable decoder model factory that produces high-performance, low-latency decoders and continually adapts them to evolving quantum hardware.
Figures & tables
Figure 1: AI model factories connect a source of training data to new models and their applications. The language-model analogy motivates a corresponding process for quantum computing, with quantum hardware providing experimentally testable outcomes. The return arrows represent model assistance to humans or quantum hardware. In the illustrated approach, data generation supplies training examples, and distillation trains later models on these examples. This paper develops the decoding instance through supervised distillation from search trajectories, with outcome verification supplied by simulation; hardware learning, control, and calibration are prospective applications.
Figure 2: Proposed coupled quantum–classical deployment of a local learned decoder. Circles represent qubits and square groups represent local classical memory. Solid and dashed links indicate quantum couplings and classical message passing; vertical arrows carry measurement data and recovery feedback between the layers.
Figure 3: Detector error model. Hidden error events connect to the syndrome bits and logical observables they flip. The decoder sees the syndrome and predicts an error pattern. Highlighted nodes illustrate one realization and its syndrome and logical outcome.
Figure 4: GraphMP predicts errors through local message passing. The syndrome enters at detector nodes; shared learned updates update node states and messages, and error nodes read out probabilities. The same learned functions operate on different decoding graphs. Hidden states and logits are illustrative.
Figure 5: BP distillation supplies a transferable starting decoder, while leaving a gap to stronger classical methods. The three shaded code-capacity tasks supply training data; the other nine evaluate the same model without fine-tuning. Logical error rates use realizations per task. Bars above are truncated. Training and comparison settings are in Appendix B .
Figure 6: Search turns successive conditional predictions into a coordinated error pattern. Each branch fixes an uncertain variable and predicts the rest again; pruning bounds the computation. The green frame marks the returned completion. This illustrative BP-guided tree shows the mechanism; the GraphMP implementation is specified in Appendix C.2 .
Figure 7: Inference-time search trades computation for accuracy. With the starting GraphMP model fixed, broader and deeper search lowers logical error while increasing decoding time. The surfaces aggregate the three training tasks; measurements and budgets are given in Appendix C.3 .
Figure 8: Outcome feedback selects search trajectories for training. The collector retains syndrome-consistent, logically correct final predictions together with the intermediate states along their search paths. This work obtains logical feedback in simulation, while dedicated quantum-memory experiments provide a route to device-based verification.
Figure 9: The GraphMP implementation alternates trajectory collection and model training. The current model guides search, selected trajectories are added to the accumulated data, and distillation produces its successor. and are the single-pass and search-enhanced decoders, and labels each trajectory dataset.
Figure 10: Successive models improve single-pass decoding and transfer that improvement to an unseen task. Dashed green curves show single-pass decoding; solid green curves show beam-search decoding. Horizontal lines show reference decoders. Curves are means over five independent model factory runs; bands are the range across those runs. Full generation results are in Appendix E.2 .
Figure 11: The final search-enhanced model compared with established decoders on tasks. The three shaded tasks supply training data; the other nine test the same model without further training. Comparisons use the same error realizations per task, except for shorter Tesseract runs on the largest tasks. Bars above are truncated. Decoder settings and evaluation procedures are specified in Appendix E.4 .
Figure 12: Accuracy and decoding time on a training task and an unseen task. Logical-error curves are fitted to measured data; bands are bootstrap intervals conditional on that model. Times compare GPU GraphMP with single-core CPU reference decoders. They measure time per shot in a stream of syndromes. Fit diagnostics and timing procedures are in Appendices E.5 and E.6 .
Figure 13: Adaptation after the noise changes at . The same model is evaluated immediately before and after the switch from depolarizing to XZ noise; – learn from newly collected trajectories. (a) supplies training data; (b) tests adaptation without further training on that task. Dashed green curves show single-pass decoding; solid green curves show beam-search decoding. Relay-BP is evaluated with its original and updated noise priors; GraphMP’s beam-search candidate selector retains the original prior. Bands are Wilson intervals. The full experiment is described in Appendix E.7 .
Appendix figures & tables29 assets
Supplementary material from the paper’s appendix.
Appendix
| Family | Instance | Training | Ref. | |
|---|---|---|---|---|
| Bivariate bicycle | yes | [ 6 ] | ||
| Bivariate bicycle | no | [ 6 ] | ||
| Generalized bicycle | yes | [ 5 ] | ||
| Generalized bicycle | no | [ 5 ] | ||
| Mitten | yes | [ 9 ] | ||
| Mitten | no | [ 9 ] |
Table 1: Code instances studied in the experiments. A yes marks a code used to train the decoder model. Distances written are the published upper bounds.
Figure 14: Detector error model. The fault is drawn from the error-variable prior. The syndrome is observed; the logical error is not.
| logical success ( ) | logical error ( ) | |
|---|---|---|
| syndrome consistent ( ) | success | miscorrection |
| syndrome inconsistent ( ) | escape | failure |
Table 2: Classification of the residual along syndrome consistency and logical outcome.
| Objective | Probability model | Decision | |
|---|---|---|---|
| MLD | |||
| , | MLE | ||
| , | bit-wise MAP |
Table 3: Three training objectives and their probability models. MLD is optimal [ 10 , 23 ] ; MLE and bit-wise MAP are successive approximations.
Figure 15: Gradient signal of the two objectives at random initialization: SNR ( 8 ) of the gradient signal on the shared model parameters, estimated over minibatches of syndromes (mean over initializations, bars the range), for qLDPC codes at code-capacity depolarizing noise . The dashed line marks . Codes sharing a distance are offset horizontally.
| Decoder | Model input representation | Prediction target |
|---|---|---|
| AlphaQubit [ 27 ] | Roundwise stabilizer observations and positional information; optional soft readout and leakage features | Logical-observable errors |
| Astra [ 28 ] | Code Tanner graph and syndrome-node features | Per-error probabilities |
| GraphQEC [ 29 ] | Syndrome history on an extended code Tanner graph with data, check, and logical nodes | Logical flips |
| Foundation decoders [ 30 ] | Transformer: detector histories and relation embeddings; Neural-BP: DEM graph, priors, and syndrome | Transformer: logical flips; Neural-BP: error mechanisms |
| GraphMP (this work) | DEM detector–error graph, syndrome, and optional pin mask | Per-error probabilities |
Table 4: Representative model interfaces. Input representations and output targets are separate design choices. The demonstrated range of parameter reuse depends on the training and evaluation protocol.
Figure 16: GraphMP forward pass. The syndrome and the mask are encoded. The decoding graph specifies the bipartite support of the check variable message-passing rounds (dashed). Readout emits logits ; per-variable probabilities are , with the logistic sigmoid Eq. ( 10 ).
| Label type | Input | Label | Teacher |
|---|---|---|---|
| Hard | simulator or a decoder | ||
| Soft | a decoder |
Table 5: Supervised formats. GraphMP distills from a simulator or a decoder teacher. A simulator supplies a hard label; a decoder may supply a hard label, a soft label, or both.
Figure 17: Logical error rate of GraphMP, its BP teacher, and reference decoders on the decoding tasks of Appendix A.1 , with samples per task. The shaded band marks the three code-capacity training tasks. The same model is applied without fine-tuning to the unshaded unseen tasks. The BP teacher, BP+OSD/LSD, and Relay-BP configurations are those of Section A.5 .
Figure 18: Decoding-outcome profiles of GraphMP in single-pass decoding, its BP teacher, and Relay-BP. Each profile classifies decoding outcomes as in Table 2 . Columns report the fraction of error realizations in each category. The logarithmic scale keeps rare outcomes visible. The shaded bands between columns follow groups of error realizations from their bucket under GraphMP to their bucket under each reference decoder; band width shows how many realizations take that path. Percentages are pooled over three code-capacity training tasks.
Figure 19: Beam search on a toy decoding task. Each box is a search state on the code’s Tanner graph. Squares are checks, filled according to the residual syndrome left by the predicted error. Circles are error variables, shaded by their predicted error probability from cream at to orange at . The soft decoder runs two rounds of belief propagation. A thick border marks a pinned variable. Each step pins the least-confident variable to and to . A state with no residual syndrome is completed (thick frame). The search returns the completed state of lowest weight, and the green edges mark the trajectory that leads to it.
| Code | Single-pass | Search ( ) | Search ( ) | |
|---|---|---|---|---|
| Pooled | — |
Table 6: Logical error rate of GraphMP , with pointwise Wilson confidence intervals. Each code is evaluated on sampled errors. Single-pass decoding and both search settings process the same samples; the pooled row contains samples. Single-pass decoding consists of one GraphMP forward pass. Search repeatedly queries that model with width and maximum depth .
Figure 20: The BP-distilled starting model at fixed search width on the three code-capacity training tasks. The left endpoint is single-pass decoding; the remaining points increase the search depth. Residual-syndrome density of the predicted error pattern is shown above logical error rate, which is decomposed into failures and miscorrections. Shaded bands are confidence intervals; shots per task, the same shots at every search depth.
Figure 21: Accuracy–cost tradeoff of beam search with the BP-distilled starting model . (a) Pooled logical error rate over search width and depth on the three training tasks at nominal evaluation error-rate settings on , on , and on ; shots per cell, with the same sampled errors in every cell. (b) Throughput-level decoding time (s/shot), averaged over the same three tasks. Circles mark the primary evaluation setting , .
| Failure rate when evaluated at | |||
|---|---|---|---|
| Model | root state | pinned states | final state |
| Root-state-trained | |||
| Trajectory-trained | |||
Table 7: The misalignment measured at the states a search visits. Two GraphMP models trained on the same trajectories, one on their root records only and one on every record, are evaluated at the recorded states of held-out successful searches on (nominal error-rate setting , search width ). The root-state-trained model fails on most pinned states even though the pinned values are correct; the trajectory-trained model turns the pinned values into completions. The pinned-state column averages over the states of a trajectory and then over trajectories. Mean standard deviation over training seeds. A prediction fails when it is syndrome-inconsistent or logically incorrect. The probe evaluates students on selected teacher states that supply correct partial information.
| Single-pass | Search ( , ) | |||
| Root-state | Trajectory | Root-state | Trajectory | |
| Joint failure rate | ||||
Table 8: The same two models as Table 7 , deployed as decoders on at nominal error-rate setting . Trajectory training improves both modes, with the larger relative gain inside beam search. A prediction fails when it is syndrome-inconsistent or logically incorrect. Every model is scored on the same error realizations; mean standard deviation over training seeds. The comparison includes differences in source sampling as well as conditional-state training.
Figure 22: The GraphMP collection-and-training loop for . Accumulated data train ; collects verified trajectories that supply records . Appending those records and advancing starts the next round. The labels B, C, and D refer to the preceding appendices.
| Generation | Initialization | Training data | Data collection |
|---|---|---|---|
| random | BP teacher | collects | |
| random | collects | ||
| collects | |||
| — |
Table 9: Model initialization, training data, and subsequent data collection. Every model trains for updates on the three target tasks. The model factory run ends at .
Figure 23: Logical error rate of GraphMP versus generation. Dashed: single-pass decoder GraphMP ; solid: search-enhanced decoder GraphMP . Horizontal lines: BP teacher, BP+OSD (BP+LSD on the phenomenological tasks), and Relay-BP. Curves are means over five independent model factory runs; bands are the range across those runs. The phenomenological panels have their own vertical axis.
| Transition | Single-pass reduction | Search-enhanced reduction |
|---|---|---|
Table 10: Reduction in the equal-task mean logical error rate, in percentage points, on the three code-capacity target tasks, , , and . Each entry is the mean reduction over five independent model factory runs, followed by the range across those runs. Every model is evaluated on the same realizations per task.
| Single-pass | Search-enhanced | |||||
|---|---|---|---|---|---|---|
| Task | Logical | Syndrome | Joint | Logical | Syndrome | Joint |
| Code-capacity | ||||||
| Code-capacity | ||||||
| Code-capacity | ||||||
| Code-capacity | ||||||
| Phenomenological | ||||||
Table 11: Final-generation logical error rate, syndrome-inconsistency rate, and joint failure rate. Entries are mean probabilities over five independently trained models, each evaluated on the same realizations per task at the settings of Figure 23 . Joint failure means that the predicted error has a wrong logical outcome, a nonzero residual syndrome, or both.
Figure 24: Later datasets challenge a fixed decoder, while later models achieve higher label-relative completion on fixed datasets. Figure 24 (a): the highest nominal error-rate setting used for collection. Figure 24 (b) and Figure 24 (c): completion relative to the recorded final label, the fraction of states on which the prediction matches that label’s syndrome and logical outcome. Each cell pools up to sampled records per target task; root and pinned states are sampled separately. Columns – are the new datasets defined in Appendix E.1 . These are diagnostics on the archived training buffers. has no pin-conditioned input and is examined only at root states.
Figure 25: Logical error rate of the trained students, single pass (top) and beam search (bottom). Each panel pools three training seeds. Filled points pool the seeds; open points are the individual seeds. The Wilson intervals summarize pooled shot-level variation; open points show variation across training runs. Scores follow Eq. ( 36 ) from the first legend entry to the second. In panel (c) both sets are collected by , one at a fixed error rate and one at an adaptive error rate. An escape counts as a logical success.
Figure 26: GraphMP against classical decoders on decoding tasks at the nominal error-rate setting under each task; the same tasks and realizations as Figure 17 , except for shorter Tesseract prefixes ( realizations on code-capacity , and , and on phenomenological ). The three shaded tasks form the model factory’s target portfolio and supply its training data. The other tasks are unseen during training and evaluate transfer beyond that portfolio. Baseline configurations are defined in Section A.5 . Bars above are truncated.
| Task | GraphMP [ CI] | BP+OSD/LSD [ CI] | Holm-adjusted paired |
|---|---|---|---|
| Code-capacity tasks | |||
Table 12: Statistical support for the fixed-model benchmark in Figure 26 . Entries give logical error probabilities and pointwise Wilson intervals, including an upper bound when no logical error is observed. BP+OSD is used on the code-capacity tasks and BP+LSD on the phenomenological tasks. Paired -values are adjusted across the comparisons shown; this family excludes the other decoder arms.
Figure 27: Logical error rate against physical error rate on the six decoding tasks of Figure 23 , for GraphMP to , BP+OSD (BP+LSD on phenomenological tasks), and Relay-BP. Curves are fitted to the measured points; bands are pointwise parametric-bootstrap intervals conditional on the fit family.
Figure 28: Decoding time per shot on the six tasks of Figure 27 . GraphMP runs on one GPU with the inference implementation described in the text (single-pass dashed, search-enhanced to solid, beam search in decoding mode); BP+OSD, BP+LSD, and Relay-BP run on one CPU core with one thread. Points above s per shot are clipped, and BP+LSD is omitted on above , where it is excluded from the accuracy evaluation as well.
Figure 29: Generation progress after the noise model changes from the depolarizing noise model to the XZ noise model at . The nominal error-rate setting and decoding graph remain fixed. Dashed: single-pass decoder GraphMP ; solid: search-enhanced decoder GraphMP ; bands are Wilson intervals on realizations per task. The BP teacher and Relay-BP step at to their logical error rate under the original prior; the faint lines give their rate under the updated prior. The model factory collects on the first three target tasks only; is unseen during adaptation.
Figure 30: Training data for noise adaptation. Starting from the same collector, equal-budget continuations train on fresh depolarizing data or fresh XZ data. Logical error rate is shown for a single pass (top) and for beam search (bottom), on the same realizations per task under each noise model, averaged over three training seeds. Wilson intervals summarize the pooled shot-level variation; the mean weights the three target tasks equally. Each panel reports of Eq. ( 36 ) from the depolarizing-data student to the XZ-data student. The unmodified mean is marked for reference. An escape counts as a logical success.