{"id":"c352a7a0-a89b-4aa2-a0e8-bc06d442b69e","arxiv_id":"2507.10876","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Multi-parallelized quantum LSTM and GRU, using one variational circuit per hidden unit with all qubits measured, achieve roughly 1.5% lower test loss than classical LSTM/GRU on the evaluated datasets.","lead":"The paper proposes a hybrid quantum-classical pipeline for forecasting high-dimensional spatial data: choose a few sensor locations, forecast them with new quantum LSTM/GRU variants, then reconstruct the full field with a decoder. It reports about 1.5% lower test loss than classical LSTM and GRU, with a 0.256% root mean squared percentage error against independent pressure measurements.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline ~1.5% test-loss gain of MP-QLSTM/MP-QGRU over LSTM/GRU rests on a single run per method, and the paper's only multi-seed experiment (Table S1) shows seed-to-seed scatter exceeding that gap, so the claimed improvement is not yet established.","rationale":"I read the central claim as having two quantitative pillars: (1) the new MP-QLSTM/MP-QGRU cells reduce test loss by about 1.5% relative to classical LSTM/GRU and about 15% relative to L-QLSTM on the Kármán vortex pressure data, and (2) the full sparse-sensing-plus-decoder pipeline reaches 0.256% RMSPE against independent pressure-tap measurements. Pillar (1) is the paper's main novelty; pillar (2) is its demonstration. For pillar (1) to hold, the reported gap must be a real effect rather than seed-to-seed variation, and that condition is the least secure. The manuscript reports one training run per method for the main experiment, and the only place it quantifies variance (Supplementary Note 4, Table S1) shows relative standard deviations of 3-5% of the MSE for both MP-QLSTM and LSTM, larger than the 1.4% gap claimed in the main experiment; moreover the Table S1 directions are inconsistent across noise levels. A reader therefore cannot currently distinguish the headline improvement from the noise floor. I partially agree with the reader's weakest assumption: using full-data POD modes (Eq. 20) to select sensor positions before the 70/20/10 split is a real protocol violation, and the autoencoder/transformer training split per Eqs. (15)-(19) is never explicitly restricted to the training period either, so the absolute 0.256% reconstruction accuracy lacks a clean generalization guarantee. However, because all five forecasting methods share the same sensors, the sensor leak biases the absolute accuracy claim but not the relative ordering of the architectures, whereas the unreplicated comparison directly threatens the paper's central architectural assertion. Credit where due: the MP-QLSTM/MP-QGRU construction is clearly specified and novel; the Lorenz experiment is multi-seed (good practice, even though its spread undermines the headline); the NREL experiment provides an independent dataset; and if 0.256% survives both a training-only sensor reselection and multi-seed reruns, it would be a solid engineering demonstration. The code-availability mismatch (the promised code is for time-series clustering, not forecasting) is a reproducibility gap rather than a correctness threat. The concrete test above settles the primary concern cheaply and directly, and it does not change the conditional verdict: the paper is promising but its headline comparison is currently unverified, so the verdict should remain CONDITIONAL pending seed-level evidence and a properly split sensor-selection protocol.","tokens_in":20409,"tokens_out":21040,"duration_ms":234670,"concrete_test":"Re-run the main pressure experiment exactly as specified in Methods (five sensor positions, hidden dimension 5, window length 4, Adam lr=0.02, batch size 128, L2 next-step loss, same epoch count) with at least 10 random seeds per model for LSTM, GRU, L-QLSTM, MP-QLSTM, and MP-QGRU, recording the test MSE of each seed. Compute the mean and a 95% bootstrap confidence interval for the paired difference MP-QLSTM minus LSTM (and MP-QGRU minus GRU, and MP-QLSTM minus L-QLSTM) on test MSE, and likewise for the per-seed RMSPE after the shared decoder. If the CI for a classical comparison includes zero, the abstract's 'approximately 1.5% lower test loss' statement is not supported and should be replaced with multi-seed means or a comparability statement; only if the CI excludes zero does the claimed improvement survive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that MP-QLSTM/MP-QGRU improve forecasting rests on unreplicated numbers from the Kármán vortex experiment: test MSE 6.40e-2 and 6.39e-2 versus 6.49e-2 (LSTM) and 6.47e-2 (GRU), and RMSPE 0.256% versus 0.264% (LSTM). The Methods describe one configuration per model with no seed count or repeat information, so each headline number is a single-run outcome. The only variance data in the manuscript are five-seed Lorenz runs in Supplementary Note 4, Table S1. There, MP-QLSTM versus LSTM MSE differences are 0.87 (std=1), 1.6 (std=3), and 1.5 (std=5) with per-run standard deviations of 0.8-1.9, i.e., relative standard deviations of 3-5% of the mean, while the main experiment's claimed gap is roughly 1.4% relative. The direction of superiority flips across noise levels (LSTM better on average at std=1 and 3, MP-QLSTM nominally better at std=5), and at std=5 the difference (29.2±1.0 vs 30.7±1.5 with n=5) is within one pooled standard deviation. Using the authors' own variance calibration, the headline 1.4% gap plausibly lies inside the run-to-run noise floor, so the abstract's first quantitative claim is not supported as reported. The ~15% advantage over L-QLSTM is a larger, more robust effect and may survive, but the comparison to classical LSTM/GRU, which the abstract emphasizes, needs seed-level evidence. The sensor-placement leak the reader identified is a genuine protocol violation (Eq. 20 computes POD modes from the full data matrix before the 70/20/10 split), but since all five methods share the same selected positions, it mainly inflates the absolute 0.256% reconstruction number rather than the relative ordering; the unreplicated single-run comparison is the more direct threat to the architectural claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-stage pipeline for forecasting high-dimensional spatial time-series data: (i) optimal sparse sensor placement via an annealing machine using POD modes, (ii) forecasting at those sensors with new quantum-classical hybrid architectures named MP-QLSTM and MP-QGRU (multiple VQCs per gate, all qubits measured), and (iii) full-field reconstruction via a learned decoder. The authors report for Kármán vortex street pressure data that MP-QLSTM and MP-QGRU achieve about 1.5% lower test MSE than classical LSTM and GRU, about 15% lower than L-QLSTM, and an RMSPE of 0.256% against independent pressure-tap measurements. Additional experiments on NREL solar power data and Lorenz-equation data with Gaussian noise are reported in the supplementary information.","tokens_in":20801,"tokens_out":4284,"duration_ms":51749,"significance":"If the reported gains are reproducible, the MP-QLSTM/MP-QGRU architecture is a meaningful contribution to hybrid quantum-classical sequence modeling for high-dimensional spatial data, and the full pipeline (sensor placement + forecasting + learned decoder) is practically interesting. The inclusion of independent pressure-sensor validation and the multi-seed Lorenz experiment in Supplementary Note 4 are strengths that partially support the methodology. However, the key headline numbers rest on single unreplicated runs and a sensor-selection protocol that uses the full dataset before the train/validation/test split, so the currently supported claims are narrower than the abstract suggests. The work would be significantly strengthened by multi-seed statistics for the main experiments and by a clean separation of training and test information in the sensor-placement and decoder stages.","major_comments":[{"comment":"The central quantitative claim that MP-QLSTM and MP-QGRU outperform classical LSTM and GRU by about 1.5% (test MSE 6.40e-2 and 6.39e-2 versus 6.49e-2 and 6.47e-2) is based on single runs: the Methods describe one configuration per model with no seed count or repeat information. The only variance data in the manuscript, Supplementary Note 4 Table S1, show per-run standard deviations of 3-5% of the mean on Lorenz MSE and a reversal of the ranking across noise levels (LSTM better at std=1 and std=3, MP-QLSTM nominally better at std=5). By the authors' own calibration, a 1.4-1.5% gap lies inside the run-to-run noise floor, so the abstract's first quantitative claim is not established as reported. Please provide multi-seed test-loss statistics for the Kármán and solar experiments, or explicitly present the headline numbers as single-run observations without a comparative accuracy claim.","section":"Results (Fig. 3) and Methods: Time-series forecasting using MP-QLSTM"},{"comment":"The sensor positions are selected from POD modes computed by the SVD of the full data matrix P in Eq. (20), which is performed before the 70/20/10 split described in the same subsection. This means the sensor-location choice uses information from the test period. In addition, the decoder in Eq. (19) is described as trained on the data with no explicit train/validation/test split, so the reconstruction evaluation (RMSPE 0.256% against independent pressure taps) likely incorporates test-period information through the decoder as well. Consequently, the RMSPE is not an unbiased evaluation of the forecasting pipeline on unseen data. Please re-run sensor selection and decoder training using only training-period data, or clearly state that the reported RMSPE measures reconstruction after training on the full record, not forecasting accuracy on truly unseen data.","section":"Methods, Data processing (Eq. 20) and Estimation of high-dimensional data (Eq. 19)"},{"comment":"The Code availability section states that 'The code for the time-series clustering developed in this study is included in Supplemental information,' but the manuscript contains no clustering method and the model/training code for MP-QLSTM and MP-QGRU is not provided. Given that the principal results are unreplicated and depend on specific circuit layouts, hyperparameters, and seeds, this makes independent verification difficult. Please either release the implemented code or provide a complete configuration ledger (random seeds, parameter counts, circuit construction, and training schedules) so that the experiments can be reproduced.","section":"Code availability"}],"minor_comments":[{"comment":"The relationship between the M linear layers in Eq. (2) and the encoding stages described in Fig. 7 is not fully explicit; the text mentions 'three outputs from the first linear layer' but does not state whether each linear layer outputs a vector of dimension equal to the qubit count and how the M encodings are combined. Please clarify the tensor shapes and the concatenation/ordering of the linear-layer outputs.","section":"Methods, Eq. (2) and Fig. 7"},{"comment":"The caption lists only LSTM, L-QLSTM, MP-QLSTM, and MP-QGRU, while the GRU result is relegated to Supplementary Note 1. Including GRU in the main figure or explicitly cross-referencing it would make the comparison easier to follow.","section":"Fig. 3 caption"},{"comment":"The affiliation of the Fujitsu authors contains the typo 'Fujistu Ltd' and should read 'Fujitsu Ltd.'","section":"Author affiliation"},{"comment":"The caption of Fig. S2 says 'Forecasting results of GRU and L-QGRU are shown,' but the main text of Note 2 is about GRU and L-QLSTM; the caption appears to contain a typo.","section":"Supplementary Note 2 caption"},{"comment":"The sentence 'even though the same learning rate and batch size were used for all methods, the proposed method shows less fluctuation' is grammatically redundant and should be reworded for clarity.","section":"Results, third paragraph"}],"recommendation":"major_revision","confidential_remarks":"The paper's core architectural idea is reasonable and the Lorenz multi-seed experiment is a welcome step, but the main experiments lack replication and the sensor-placement/decoder protocol conflates training and test information. These issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection. The authors should also be asked to reconcile the Code availability statement, which currently does not describe the actual code for the proposed method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the multi-parallelized idea — one VQC per cell-state dimension, all qubits measured — is a real architectural extension of QLSTM/L-QLSTM, and the paper is honest about the modest size of the gain. But the abstract's central claim, roughly 1.5% lower test loss than classical LSTM/GRU, rests on a single run per method. The only multi-seed experiment in the supplement shows run-to-run scatter larger than that gap. That needs fixing before the claim can be trusted.\n\nWhat is genuinely new: the MP-QLSTM/MP-QGRU architecture is a sensible response to the information bottleneck in earlier QLSTM variants. The paper evaluates it on three datasets, including a Kármán vortex street experiment with independent pressure-tap validation, which is a nice touch. The writing is clear, the flow data is on Zenodo, and the sensor placement step builds on prior work with an annealing machine. The comparison against L-QLSTM, where the gap is around 15%, looks more robust and likely survives replication.\n\nSoft spots, in proportion: the headline 1.5% versus LSTM/GRU is unreplicated. The Methods specify one configuration per model with no mention of seeds. The only variance data are the five-seed Lorenz runs in Table S1, where MP-QLSTM is worse at noise levels 1 and 3, and even at noise level 5 the difference (29.2±1.0 vs 30.7±1.5) sits inside the pooled spread. So the primary quantitative claim is not established as reported. Also, sensor placement uses POD modes from the full data matrix (Eq. 20) before the 70/20/10 split, leaking test-period information into the reconstruction evaluation. Since all methods share the same sensors, this mainly inflates the absolute 0.256% reconstruction number rather than the relative ordering, but it is still a protocol violation. The code availability statement says 'time-series clustering' code, not the forecasting code, which is misleading.\n\nWho this is for: people working on quantum RNNs or hybrid quantum/classical forecasting will want to know this architecture. It is a plausible increment, not a breakthrough. I would send it to peer review, but with a requirement for seed-level results on the main comparison and a corrected sensor-selection protocol before acceptance. The architecture deserves discussion; the empirical claim as stated does not yet deserve the headline.","headline":"The MP-QLSTM/MP-QGRU architecture is a genuine variant worth knowing about, but the headline 1.5% edge over LSTM/GRU is not yet supported by the evidence.","tokens_in":21423,"tokens_out":1785,"would_cite":true,"duration_ms":21067,"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":"MP-QLSTM and MP-QGRU — recurrent networks in which every cell-state dimension owns a variational quantum circuit with all qubits measured — achieve roughly 1.5% lower test loss than LSTM and GRU, about 15% lower than L-QLSTM, and 0.256%…","keywords":["time-series forecasting","high-dimensional spatial data","variational quantum circuits","quantum long short-term memory","gated recurrent unit","optimal sensor placement","autoencoder decoder","pressure-sensitive paint"],"falsifier":"Re-run the full pipeline with sensor placement computed only from the training portion of the PSP dataset, then report test MSE and RMSPE on the untouched test portion; if the 0.256% RMSPE and the 1.5% advantage over LSTM shrink or vanish, the reported accuracy depends on test-period information rather than on MP-QLSTM alone. A second check is to replace each variational circuit with a classical nonlinear layer of matched parameter count and see whether the claimed improvement persists.","tokens_in":20198,"feed_emoji":"🌀","tokens_out":6131,"duration_ms":64968,"temperature":0.7,"pith_summary":"This paper proposes an end-to-end recipe for forecasting high-dimensional spatial time series: choose a handful of sensor locations that best represent the field, run a recurrent forecast at those points, and reconstruct the full field with a trained autoencoder decoder. The forecasting step introduces two quantum recurrent architectures, MP-QLSTM and MP-QGRU, that give each cell-state dimension its own variational quantum circuit and measure all qubits, in contrast to earlier QLSTM designs that discard part of the circuit's output. On the pressure field behind a square cylinder (Kármán vortex street), the quantum hybrids achieve about 1.5% lower test loss than LSTM and GRU and about 15% lower than the linear-layer-enhanced L-QLSTM, and MP-QLSTM reconstructs the full field with 0.256% root mean squared percentage error against independent pressure-tap measurements. The same architecture also improves solar power forecasting by roughly 3% over classical baselines, and the paper argues the method is most useful when short-term dependencies dominate and measurement records are limited.","feed_headline":"Hybrid quantum network tops LSTM by 1.5% on pressure forecast","feed_subtitle":"Five smartly placed sensors feed MP-QLSTM; a decoder rebuilds the full 780×780 flow field at 0.256% error.","key_machinery":"The central object is the multi-parallelized variational quantum gate. For each gate (forget, input, cell, output), the hidden state and input are concatenated, then M linear layers project the combined vector to M values, each encoded as rotation angles on one qubit via RX, RY, and RZ gates; CNOT entangling layers are interleaved with rotations; every qubit is measured; and the K circuit outputs form the gate vector. The same number K of circuits equals the cell-state dimension, and this full measurement is what gives the architecture its claimed extra representation capacity.","core_discovery":"The central claim is that the information bottleneck in quantum LSTM is not the number of qubits but how the circuit's outputs are used. Original QLSTM maps the concatenated input and hidden state into one variational quantum circuit and measures only a subset of qubits to match cell-state dimension; L-QLSTM pushes all inputs through a linear layer, losing information before encoding. MP-QLSTM instead runs K separate variational quantum circuits, one per cell-state dimension, each fed by its own M linear projections so every qubit carries usable data, and all qubits are measured. This parallel structure is a strict generalization: with M=1 and an identity circuit it reduces to classical LSTM. The paper reports that this design lowers test MSE from 6.47e−2 and 6.49e−2 (GRU and LSTM) to 6.39e−2 and 6.40e−2 (MP-QGRU and MP-QLSTM) on the flow data, and gives a 0.256% RMSPE against independent pressure sensor measurements after full-field reconstruction.","pith_inferences":["The paper does not isolate whether the gain comes from quantum expressivity or simply from a larger number of trainable parameters; a fair control is a classical network with K×M parallel nonlinear layers replacing each variational circuit.","Using POD modes from the full dataset for sensor placement is a form of transductive learning; a train-only placement would be the deployment-realistic version and the most direct stress test of the 0.256% figure.","The pipeline invites extension to other reduced-order bases such as dynamic mode decomposition, and to online sensor selection where positions adapt as new measurements arrive.","The reported comparisons use simulated circuits; running the same circuits on noisy quantum hardware would show whether the advantage survives measurement and gate noise."],"forward_implications":["If full-qubit measurement is the cause of the gain, future QLSTM variants should measure all qubits rather than matching cell-state dimension by discarding measurements.","The three-step pipeline (optimal sensor placement, sparse forecast, decoder reconstruction) makes high-dimensional 780×780 flow forecasting feasible with as few as five physical sensors and 0.256% error.","On the solar power benchmark, the same architecture transfers to a different nonlinear domain with a roughly 3% MSE improvement over LSTM and GRU.","In noisy Lorenz data, MP-QLSTM beats LSTM at high noise levels while LSTM remains better at low noise, suggesting a data-dependent choice of method.","Because MP-QLSTM reduces to classical LSTM when the circuit is the identity and M=1, any reported improvement is a strict extension rather than an unrelated model class."],"supporting_citations":[{"why":"Introduces QLSTM, the variational-circuit-in-LSTM architecture that MP-QLSTM extends by using one circuit per cell-state dimension and measuring all qubits.","marker":"76"},{"why":"Proposes L-QLSTM, the linear-layer-enhanced baseline whose test loss MP-QLSTM and MP-QGRU claim to beat by about 15%.","marker":"78"},{"why":"Supplies the annealing-machine optimal sensor placement method and the Kármán vortex street dataset used for the main demonstration.","marker":"107"},{"why":"Provides the pressure-sensitive paint measurement data for the Kármán vortex street flow field used in training and evaluation.","marker":"108"},{"why":"Is the earlier data-driven noise reduction work whose truncated-SVD approach is used here to denoise the pressure data.","marker":"109"},{"why":"Provides the solar power dataset used as a second benchmark for MP-QLSTM and MP-QGRU.","marker":"110"},{"why":"Supplies the multi-task learning formulation that jointly trains the encoder, decoder, and sensor-to-latent transformer in Eq. (19).","marker":"111"}],"fun_headline_variants":["Quantum LSTM with full qubit readout beats classical by 1.5%","Full-qubit measurement in quantum recurrent nets improves LSTM","Parallel quantum circuits measure all qubits, cut loss 1.5%","MP-QLSTM's full qubit readout yields 1.5% gain over LSTM","Quantum recurrent unit with full measurement outperforms LSTM"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Sensor positions are chosen by solving the optimal sensor placement problem on POD modes computed from the entire pressure data matrix, before the data are split into train, validation, and test sets, so test-period information can influence where the five sensors sit; if sensor placement is restricted to training data only, the reported forecasting and reconstruction errors may be optimistic.","fun_headline_variants_meta":{"raw":{"variants":["Quantum LSTM with full qubit readout beats classical by 1.5%","Full-qubit measurement in quantum recurrent nets improves LSTM","Parallel quantum circuits measure all qubits, cut loss 1.5%","MP-QLSTM's full qubit readout yields 1.5% gain over LSTM","Quantum recurrent unit with full measurement outperforms LSTM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000678,"raw_usage":{"total_tokens":3100,"prompt_tokens":978,"completion_tokens":2122,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":2023}},"tokens_in":594,"tokens_out":2122,"duration_ms":18431,"temperature":1.0,"reasoning_tokens":2023,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:22:32.243140+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the full pipeline with sensor placement computed only from the training portion of the PSP dataset, then report test MSE and RMSPE on the untouched test portion; if the 0.256% RMSPE and the 1.5% advantage over LSTM shrink or vanish, the reported accuracy depends on test-period information rather than on MP-QLSTM alone. A second check is to replace each variational circuit with a classical nonlinear layer of matched parameter count and see whether the claimed improvement persists.","supporting_citations":[],"review_version":1}