{"id":"94ede809-d926-4145-9a82-7646f9e244ea","arxiv_id":"2504.16999","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A modular neural decoder, MCCD, learns to decode surface-code logical circuits with single-qubit and entangling gates, keeping accuracy while running faster than MWPM, MLE, and BP-OSD in simulated benchmarks.","lead":"This paper introduces a modular machine-learning decoder, MCCD, that learns to decode logical quantum circuits with both single-qubit and entangling gates, trained on mirror-symmetric random Clifford circuits. It reports competitive accuracy and faster wall-time than conventional decoders in simulations, but only for one noise model and two circuit templates.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training and test are both mirror-symmetric, so the claimed generalization to 'logical circuits' may be an artifact of the mirrored training distribution; non-mirror circuits are never tested.","rationale":"The reader's conditional verdict is appropriate; I do not move it. My review agrees with the core concern about distribution shift but sharpens it to a property of the training scheme itself. The abstract's claim of a noise-model-agnostic decoder for logical circuits rests on experiments where both train and test circuits are mirror-symmetric random Clifford circuits; only depth and distance vary. An LSTM with memory could exploit the forward/reverse symmetry (e.g., by using reverse-half syndromes to revisit errors from the forward half), so the high accuracy in Figs. 4-5 may not indicate general circuit-decoding ability. A direct non-mirror test is cheap with stim because logical labels can be computed by comparing noisy and noiseless stabilizer simulations. I credit the authors for releasing code and using a realistic circuit-level noise model, and the paper does not claim machine-checked proofs; but the missing non-mirror benchmark is a concrete, addressable gap. If the non-mirror test passes, the concern is resolved and the conditional can be lifted; if it fails, the central claim must be narrowed.","tokens_in":27701,"tokens_out":6051,"duration_ms":61909,"concrete_test":"Generate non-mirror random Clifford circuits of the same gate set and depths (e.g., D=4,8,12,16,20 for Type I and D=4,8,12,16,20,36 for Type II), with no reverse half, and obtain ground-truth logical-error labels by comparing the noisy final measurement with the ideal noiseless output computed by stabilizer simulation in stim. Run the already-trained MCCD (trained only on mirror circuits) on these circuits and compare per-qubit logical accuracy and wall time with the mirror-circuit results in Figs. 4-5. If accuracy is statistically equivalent, mirror symmetry is not the source of generalization; if accuracy drops materially (e.g., by more than the error-bar scale), the central claim must be restricted to mirrored benchmarking circuits or the model must be retrained on non-mirror labels.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The single most load-bearing gap is that every training and test circuit is mirror-symmetric: a random Clifford circuit U followed by its reverse U†. This is not just a labeling convenience; it gives the syndrome stream a special temporal structure. Errors created in the forward half are re-encountered in reverse order and conjugated by gates in the backward half, so an LSTM with memory has an extra source of correlation to exploit that is absent in a general logical algorithm. The paper's only held-out variation is circuit depth (and d=3 vs d=5); non-mirror circuits, arbitrary gate sequences, or circuits with multiple CNOTs per qubit per layer are never tested. Thus the demonstrated 'ability to learn correlated decoding patterns' may be specific to mirrored circuits, and the abstract's 'general circuit' and 'noise-model agnostic' claims outrun the evidence. The architecture itself does not explicitly condition on mirror symmetry, so the risk is that the training distribution, not the decoder, is doing the work. This is a correctness risk because if MCCD fails on non-mirror circuits, the central claim of a general logical-circuit decoder fails.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces MCCD, a modular LSTM-based decoder for logical circuits built from transversal Clifford gates on surface-code logical qubits. The decoder maintains one hidden state per logical qubit, processes syndromes with gate-specific recurrent cells, and uses a two-qubit cell for CNOT-induced correlated errors. The authors train on mirror-symmetric random Clifford circuits (single-qubit Type I and CNOT-containing Type II) generated with Stim under a circuit-level noise model motivated by neutral-atom hardware, then test on unseen circuits of greater depth and at code distances d=3 and d=5. They benchmark logical accuracy and wall time against MWPM, MLE, and BP-OSD, reporting that MCCD achieves accuracy competitive with BP-OSD and superior speed scaling with circuit depth, especially for entangling circuits. The paper claims the framework is a noise-model-agnostic, general logical-circuit decoder.","tokens_in":27989,"tokens_out":2400,"duration_ms":25736,"significance":"If the central claims hold, the paper makes a useful contribution: a modular recurrent decoder that processes syndrome streams with per-gate cells is a natural architecture for logical circuits, and the ability to decode CNOT-induced correlated errors with roughly linear depth scaling would be practically valuable. The manuscript has real strengths: it provides open code, uses Stim for reproducible simulation, benchmarks against three established decoders, and demonstrates depth extrapolation beyond training depths. However, the empirical evidence is narrower than the abstract suggests: only mirror-symmetric circuits are tested, only one noise model is used, and the timing/accuracy figures lack statistical error bars. With additional tests and more careful claims, the work could be an important step toward practical circuit decoding; in its current form, the generality claims outrun the data.","major_comments":[{"comment":"All training and test circuits are mirror-symmetric: a random unitary U followed by its reverse U†. This gives the syndrome stream a special temporal structure: errors created in the forward half are re-encountered in reverse order and conjugated by the backward gates, which an LSTM with memory can exploit. The only held-out variation is circuit depth (and d=3 vs d=5); non-mirror circuits, arbitrary gate sequences, and circuits with more than one CNOT per qubit per layer are never tested. The claim of a 'general circuit decoder' therefore is not established. I request experiments on non-mirror circuits (e.g., random Clifford circuits without enforcing U†, or deterministic logical algorithms such as Bell-state preparation followed by verification), plus circuits violating the one-CNOT-per-qubit-per-layer restriction, to show that the decoder generalizes beyond the mirrored training distribution.","section":"Main text, 'Mirror-symmetric random circuits' and Figs. 3-5"},{"comment":"The abstract claims a 'noise-model agnostic solution,' but MCCD is trained and tested on trajectories generated from a single circuit-level noise model (SM C.2), and the conventional decoders are given that exact same noise model. No experiment varies the noise parameters or trains on one noise model and tests on another. As written, the noise-model-agnosticism claim is unsupported. Either add cross-noise-model transfer tests (e.g., train under the neutral-atom-inspired model and evaluate under depolarizing noise with different error rates, or at least under a perturbed noise channel) or substantially weaken the claim to 'does not require an explicit noise model at inference time.'","section":"Abstract and SM C.2"},{"comment":"The central quantitative claims—logical accuracy and wall-time scaling—are presented without error bars, confidence intervals, or even the number of test trajectories used for each point. This matters because the differences between MCCD and BP-OSD accuracy appear small in some regimes, and the wall-time comparisons depend on implementation and hardware, which are not specified. Please report the sample size per data point, statistical uncertainty in the logical error rates, and the CPU/GPU hardware and software versions used for the timing measurements, and ideally run repeated timing measurements to estimate variability.","section":"Figures 4 and 5"}],"minor_comments":[{"comment":"The first sentence contains a grammatical error: 'an efficient polynomial-time decoding algorithms' should be 'efficient polynomial-time decoding algorithms.'","section":"Abstract"},{"comment":"'correlate errors' should be 'correlated errors' in the sentence 'We find MCCD to learn to decode correlate errors from a general circuit successfully.'","section":"Main text, paragraph after Fig. 1"},{"comment":"The LSTM equations use h_i for the hidden state at time i, but the surrounding text refers to h_t and h_{t-1}; aligning notation would improve readability.","section":"SM A, Eq. (A1)"},{"comment":"The table entry 'Training data size 500,000' combined with 'batch size 1024' does not immediately convey that this means 5×10^8 trajectory samples; the text clarifies this, but a footnote in the table would help.","section":"SM B, Table III"},{"comment":"The sentence 'Note - Toward the conclusion of this work, we noticed a recent work...' is awkward and sounds like an acknowledgment rather than a scientific note. It would be clearer as a normal reference comparison in the introduction or related-work section.","section":"Main text, 'Note' paragraph"}],"recommendation":"major_revision","confidential_remarks":"The refereeing process highlights a genuine tension: the paper is well-executed within its chosen setup, but the abstract and conclusion claim generality ('logical circuits', 'noise-model agnostic') that the experiments do not support. The mirror-symmetry issue is the most serious because it affects the interpretation of every reported success. I believe the authors can address this by adding non-mirror circuit tests and by retracting or qualifying the generality claims; this is why I recommend major revision rather than rejection. I would also encourage the editor to ensure that the timing benchmarks are reproducible by requesting the exact hardware/software environment as part of the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a decent incremental ML decoder paper with a real architectural idea, but the headline claim that it decodes \"logical circuits\" outruns the evidence. The authors only ever train and test on mirror-symmetric circuits, and that is a bigger gap than the depth generalization they advertise.\n\nWhat's new: the modular per-gate LSTM design, with a dedicated two-qubit cell for CNOT, is a clean way to handle correlated errors from entangling gates. Prior neural decoders treated a single logical qubit or single-qubit logical operations (Ref [3] does single-qubit Clifford circuits on the color code). The two-circuit benchmark, Type I single-qubit and Type II with logical CNOTs, is sensible, and the curriculum training—single-qubit modules first, then freeze them and train only the CNOT cell—is a practical training trick. The wall-time scaling, roughly linear in depth, is useful if it holds. They are also honest about Ref [3]'s proximity, which is refreshing.\n\nThe soft spots are real. The most load-bearing: every training and test circuit is mirror-symmetric (random U followed by U†). That gives the syndrome stream a very particular temporal structure—errors created in the forward half are re-encountered in reverse order, conjugated by the backward gates. An LSTM with memory can exploit that. Non-mirror circuits, arbitrary gate sequences, multiple CNOTs per qubit per layer—none are tested. So the abstract's \"noise-model agnostic\" and general logical circuit claims are not supported. The accuracy figures have no error bars; the timing hardware isn't specified; and the noise model is fixed (their neutral-atom motivated model). The trained model is also evaluated on the same simulator and noise model it was trained on, so part of the performance is in-distribution fitting. That is not disqualifying—the comparisons to MLE, MWPM, and BP-OSD are legitimate—but it is a reason to read the generality claims with caution.\n\nWho this is for: people working on ML decoders for fault-tolerant circuits will find the architecture worth knowing; people evaluating decoder methods for real hardware should wait for tests on non-mirrored circuits and other noise models. It deserves a serious referee: the work is clear, reproducible in principle, and the central limitation can be fixed with an additional test section.","headline":"Useful modular LSTM decoder for logical circuits, but the mirror-symmetric-only evaluation undercuts the general-circuit claims.","tokens_in":28468,"tokens_out":2500,"would_cite":true,"duration_ms":23411,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a modular recurrent decoder with one processing cell per logical gate can decode deep logical Clifford circuits with correlated errors from transversal CNOT gates, achieving accuracy competitive with exhaustive…","keywords":["quantum error correction","logical circuit decoding","correlated errors","transversal CNOT","surface code","recurrent neural network decoder","LSTM","mirror-symmetric random Clifford circuits"],"falsifier":"A concrete test: take the trained MCCD and run it on logical circuits that are not mirror-symmetric and whose physical noise is strongly biased, for example dominated by dephasing rather than the mixed Pauli channels simulated here, holding code distance and error rate fixed; if logical accuracy falls well below most-likely-error decoding or belief-propagation with ordered-statistics post-processing, or if wall time becomes superlinear with depth, the paper's generalization and scaling claims would be contradicted.","tokens_in":27536,"feed_emoji":"⚛️","tokens_out":8030,"duration_ms":68204,"temperature":0.7,"pith_summary":"Quantum error correction keeps a logical qubit alive, but decoding the syndrome data fast enough for real computation is a bottleneck, and entangling gates such as CNOT make it worse: a physical error on one qubit can propagate to another, producing correlated syndromes that standard matching decoders cannot handle directly. The paper argues that a single data-driven decoder, MCCD, can learn to handle these correlated patterns without being handed a noise model. MCCD is a modular recurrent network with one LSTM processing cell per logical gate and one hidden state per logical qubit; the CNOT cell ingests the hidden states and syndromes of both the control and target qubit, so error propagation is represented inside the network rather than in a hand-built graph. Trained on mirror-symmetric random Clifford circuits, where the forward and inverse circuits cancel so the correct answer is known, MCCD reports logical accuracy competitive with exhaustive most-likely-error decoding and with belief-propagation decoders, while its wall time grows roughly linearly with circuit depth for surface-code distances 3 and 5. If the claim holds, it removes a classical bottleneck for fault-tolerant logical circuits.","feed_headline":"A modular neural decoder learns correlated circuit errors","feed_subtitle":"Trained on mirror-symmetric random Clifford circuits, it stays accurate beyond training depth and runs in near-linear time.","key_machinery":"The load-bearing object is a set of gate-specific LSTM processing cells, meaning long short-term memory recurrent units, connected through per-qubit hidden states, together with the mirror-symmetric random Clifford circuit used for training. A processing cell for a single-qubit gate updates one qubit's hidden state from the new syndrome; the two-qubit CNOT cell concatenates the hidden states and syndromes of control and target, updates both together, and splits the result back. Mirror-symmetric circuits supply ground-truth labels with no extra simulation cost, because a noiseless forward-then-inverse circuit returns the initial state; this lets the decoder be trained end to end with a cross-entropy loss while randomizing over circuits to avoid overfitting. The two-stage curriculum, which trains single-qubit modules first and then freezes them to train only the CNOT module, keeps training data cheap and focuses capacity on the correlated-error structure.","core_discovery":"The central claim is that decoding a logical circuit can be recast as a supervised per-qubit classification problem and solved by a modular recurrent network whose structure mirrors the circuit's gate set. Each logical qubit carries a hidden state that tracks its error history, and each supported gate, including the entangling CNOT, has a dedicated processing cell that updates the involved qubits' states from the incoming syndrome. The CNOT cell receives the hidden states and syndromes of both control and target qubits at once, so it can represent the correlated errors that arise when a physical error on one qubit propagates through the gate to the other. The paper reports that MCCD, trained with a curriculum that first learns single-qubit gates and then adds the two-qubit cell, achieves logical accuracy competitive with most-likely-error decoding and belief-propagation with ordered-statistics post-processing on surface-code logical qubits of distance 3 and 5, while its wall time grows approximately linearly with circuit depth and remains far below the exhaustive methods.","pith_inferences":["Beyond the paper: if the linear wall-time scaling persists at larger code distances and with denser entangling layers, MCCD-type decoders could serve as real-time decoders inside the control loop of a fault-tolerant processor, not just as offline verification tools.","Beyond the paper: the mirror-symmetric training trick is not tied to Clifford gates; the same circuit-equals-its-inverse label scheme could generate training data for teleported non-Clifford gates, although the paper only tests transversal Clifford operations.","Beyond the paper: the noise-model-agnostic claim is only as broad as the single simulated neutral-atom-style noise model tested; a natural check is to retrain on a depolarizing or biased-noise model and compare transfer, a test the architecture invites but the paper does not perform.","Beyond the paper: the per-qubit LSTM hidden state has a fixed memory window, so very long-range temporal correlations in deep circuits may eventually be missed; extending the cells with attention or larger context would be a measurable test of that limit."],"forward_implications":["MCCD generalizes to logical circuits significantly deeper than any seen in training, with no observed degradation in logical accuracy for the tested depths.","Its wall time grows approximately linearly with circuit depth for code distances 3 and 5, making it markedly faster than exhaustive most-likely-error decoding and than belief-propagation with ordered-statistics post-processing at large depth for distance 5.","A single unified framework handles both single-qubit and entangling logical gates, including the CNOT-induced correlated errors that matching-based decoders cannot capture directly.","Because it learns from syndrome data rather than from a hand-built noise graph, MCCD is noise-model agnostic and only assumes the hardware's native logical gate set.","The modular design extends to new gate sets as hardware evolves, since each new logical operation gets its own processing cell."],"supporting_citations":[{"why":"supplies the LSTM-based recurrent decoding approach that MCCD extends from quantum memory to logical circuits.","marker":"[20]"},{"why":"provides the correlated-decoding most-likely-error baseline and the simulation approach for errors propagating through transversal gates.","marker":"[14]"},{"why":"motivates the neutral-atom experimental setting and the circuit-level noise model used to generate the data.","marker":"[2]"},{"why":"logical randomized benchmarking, whose random circuits inspire the mirror-symmetric training scheme that yields ground-truth labels.","marker":"[26]"},{"why":"the fast stabilizer-circuit simulator used to generate all training and test syndrome trajectories.","marker":"[27]"},{"why":"the minimum-weight perfect matching decoder that serves as a speed and accuracy baseline.","marker":"[12]"},{"why":"the belief-propagation with ordered-statistics post-processing decoder used as a comparison baseline.","marker":"[30]"},{"why":"the long short-term memory cell that is the building block of every gate-specific processing cell.","marker":"[24]"}],"fun_headline_variants":["Modular decoder learns correlated circuit errors","MCCD: data-centric decoder for logical quantum circuits","Neural decoder beats MWPM on deep logical circuits","Scalable modular decoder for error-corrected circuits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy and speed results transfer from the two mirror-symmetric random Clifford circuit templates and the single circuit-level noise model used in training to the logical circuits and hardware noise encountered in deployment.","fun_headline_variants_meta":{"raw":{"variants":["Modular decoder learns correlated circuit errors","MCCD: data-centric decoder for logical quantum circuits","Neural decoder beats MWPM on deep logical circuits","Scalable modular decoder for error-corrected circuits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1468,"prompt_tokens":958,"completion_tokens":510,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":449}},"tokens_in":574,"tokens_out":510,"duration_ms":4882,"temperature":1.0,"reasoning_tokens":449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:51:54.429415+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: take the trained MCCD and run it on logical circuits that are not mirror-symmetric and whose physical noise is strongly biased, for example dominated by dephasing rather than the mixed Pauli channels simulated here, holding code distance and error rate fixed; if logical accuracy falls well below most-likely-error decoding or belief-propagation with ordered-statistics post-processing, or if wall time becomes superlinear with depth, the paper's generalization and scaling claims would be contradicted.","supporting_citations":[{"cited_title":"Machine-learning-assisted correction of correlated qubit errors in a topological code","cited_arxiv_id":"1705.07855","evidence_quote":"supplies the LSTM-based recurrent decoding approach that MCCD extends from quantum memory to logical circuits."},{"cited_title":"Higgott, PyMatching: A Python Package for De- coding Quantum Codes with Minimum-Weight Perfect Matching, ACM Transactions on Quantum Computing 3, 1 (2022)","cited_arxiv_id":null,"evidence_quote":"the minimum-weight perfect matching decoder that serves as a speed and accuracy baseline."}],"review_version":1}