{"id":"4993b2f1-fff3-4282-aca1-e1da79099223","arxiv_id":"2412.08851","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A quantum kernel-based LSTM reports 42% lower RMSE than a classical LSTM on one AQI forecasting benchmark, with far fewer parameters.","lead":"This paper inserts quantum kernel evaluations into LSTM memory cells and tests the hybrid on air quality forecasting for Bengaluru, India. It reports lower prediction errors than a classical LSTM with fewer trainable parameters, though the comparison rests on a single run with no error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eqs. (6a)-(6c) compute scalar gate values, yet (6d) needs vectors of hidden dimension 16; the reported QK-LSTM cannot be implemented as written, and the 209-parameter count is inconsistent with the stated equations.","rationale":"The reader's weakest_assumption correctly identifies the scalar/vector inconsistency in the gate equations as load-bearing. My independent reading confirms that Eqs. (6a)-(6c) produce scalars, while the LSTM recurrences (6d)-(6f) require vectors of dimension m. This is an internal inconsistency, not merely an underspecification that could be resolved by a plausible default: no natural broadcasting rule recovers a 16-dimensional cell state, and the paper gives no vector-valued kernel construction. The parameter-count arithmetic (4N+4 = 209 with no integer N) provides an independent, concrete contradiction to the reported compression claim. I also note the qubit-encoding mismatch (4 qubits cannot encode the 27-dimensional concatenated vector v_t), which further undermines implementability. In good faith, the paper's aim is to propose a quantum-kernel-based LSTM and demonstrate empirical gains; however, the central empirical comparison depends on a concrete, trainable architecture that does not exist as written. The lack of code/data and the absence of a classical kernel baseline are additional reasons the empirical claim cannot be validated, but the architectural inconsistency alone suffices to reject. Since the reader's verdict already reflects this, no verdict change is needed.","tokens_in":10869,"tokens_out":2854,"duration_ms":33539,"concrete_test":"Write out the tensor shapes for Eqs. (6a)-(6d) with m = 16 hidden units and input dimension 11. Attempt a forward pass for a single time step using arbitrary values for N, reference vectors, α, and biases. If f_t, i_t, C_tilde_t, and o_t are computed exactly as scalar weighted sums, the elementwise multiplications in (6d) and (6f) will fail unless the scalars are broadcast, in which case the hidden and cell states are 1-dimensional, contradicting the stated hidden size. Additionally, solve 4N + 4 = 209 for an integer N; if no integer exists, the parameter count in Table II is inconsistent with the model definition. This analytical test settles whether the architecture is implementable as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that QK-LSTM improves AQI forecasting accuracy while reducing parameters, but the model definition in Section III-B2 is internally inconsistent. Equations (6a), (6b), (6c), and (6e) each define a gate output as a weighted sum of scalar quantum kernels k(v_t, v_j) plus a bias, followed by an activation. Each of f_t, i_t, C_tilde_t, and o_t is therefore a scalar in R. However, Eq. (6d) requires C_t = f_t ⊙ C_{t-1} + i_t ⊙ C_tilde_t, where C_{t-1} and C_tilde_t are stated to be hidden-state vectors of dimension m = 16 (Table II). Scalar-vector elementwise multiplication is undefined unless the scalars are broadcast, but broadcasting would force the cell state to be one-dimensional, contradicting the reported hidden size. The paper never specifies how the scalar kernel expansion is lifted to vector gates, how many reference vectors N are used, how they are selected, or how each hidden unit obtains a distinct gate value. This is not a minor omission: as written, the forward pass cannot be executed for the stated configuration. The parameter count compounds the problem. With only the α weights and biases described in the training section, the total trainable parameters is 4N + 4 (four gates, each with N kernel weights and one bias). Setting 4N + 4 = 209 yields N = 51.25, which is not an integer. The circuits in Eqs. (9)-(12) have no trainable parameters, since θ_k and φ_k are functions of the input v, and the parameter-shift rule in Eq. (15) is applied to parameters that do not exist in the defined circuit. Thus the reported parameter reduction is not derivable from the model as specified. A further dimension mismatch is also present: v_t = [h_{t-1}; x_t] has dimension n + m, while the quantum circuit uses 4 qubits and encodes only 2n_qubits = 8 angles, so a 27-dimensional input cannot be encoded by the stated U(v). These inconsistencies make the empirical comparison in Table I unsupported by a well-defined model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes QK-LSTM, a hybrid quantum-classical recurrent architecture in which the linear transformations inside an LSTM cell are replaced by weighted sums of quantum kernel evaluations. The model is applied to daily Air Quality Index (AQI) forecasting for Bengaluru, India, using eleven selected pollutant features. The authors report that QK-LSTM outperforms a classical LSTM on RMSE (9.20 vs 15.94), MAE (7.15 vs 11.07), MAPE (9.14% vs 13.32%), and R² (0.84 vs 0.78), while reducing trainable parameters from 1,873 to 209. The paper also discusses scalability and NISQ-era practicality, including GPU-based classical simulation of the quantum kernels.","tokens_in":11334,"tokens_out":3751,"duration_ms":43164,"significance":"If the central claims were substantiated, the paper would offer a useful data point for hybrid quantum-classical sequence modeling, particularly in showing that kernel-based replacements for learned linear projections can reduce parameter counts without losing accuracy on a real forecasting task. The authors are also to be credited for using standard evaluation metrics, documenting the data preprocessing pipeline, and clearly stating the LSTM baseline equations. However, the central contribution is not currently assessable: the defining equations do not specify a constructible model, the reported parameter count is inconsistent with those equations, and the empirical comparison rests on a single run with no error bars, no described split, and no classical kernel baseline. The paper also contains no code or data release that would allow the reported numbers to be reproduced. The significance of the claimed result therefore cannot be established from the manuscript as written.","major_comments":[{"comment":"The gate equations define each of f_t, i_t, C_tilde_t, and o_t as a scalar, because each is a weighted sum of the scalar kernels k(v_t, v_j) plus a bias, followed by an activation. Equation (6d) then requires elementwise multiplication of these scalars with the hidden-state vectors C_{t-1} and C_tilde_t, which are stated to have dimension m = 16 in Table II. The paper never specifies how the scalar gate values are lifted to 16-dimensional gate vectors, how many reference vectors N are used, how they are selected, or whether they are trainable. As written, the forward pass cannot be executed for the reported configuration. In addition, Eq. (5) concatenates h_{t-1} (dimension 16) with x_t (dimension 11) to form v_t of dimension 27, while the quantum circuit in Eqs. (9)-(12) uses n = 4 qubits. The encoding map U_enc in Eq. (10) has only 2n = 8 rotation angles for the whole 27-dimensional input, and no dimensionality reduction or component-selection procedure is described. This is a load-bearing gap: the reported architecture is not implementable as stated.","section":"III-B2, Eqs. (6a)-(6f)"},{"comment":"The reported parameter count is inconsistent with the stated model. If the only trainable parameters are the alpha weights and biases in Eqs. (6a)-(6e), the total is 4N + 4 (four gates, each with N kernel weights and one bias). Setting 4N + 4 = 209 gives N = 51.25, which is not an integer. If, alternatively, the quantum circuit parameters are trainable, then the paper must say so and explain the parameter-shift rule of Eq. (15) in that context, but Eqs. (9)-(12) define theta_k and phi_k as functions of the input v, with no free variational parameters. The parameter-shift rule is therefore not applicable to any circuit parameter described in the paper. The claimed 209-parameter model cannot be reconciled with the equations, and the model-compression comparison in Table II is not verifiable.","section":"III-B4 and Table II"},{"comment":"The empirical evaluation is not sufficient to support the central claim that QK-LSTM outperforms LSTM on AQI forecasting. The paper reports a single set of metrics with no error bars, no number of random seeds, and no description of the train/validation/test split. It also does not state whether the linear interpolation in Section II-C was performed before or after splitting the time series, which matters for data leakage and for the validity of the reported test performance. Furthermore, the comparison includes only a plain LSTM; there is no classical kernel-based LSTM baseline (for example, an LSTM whose gates use a classical RBF kernel expansion), so the reported improvement cannot be attributed to the quantum feature map rather than to the kernel expansion or to the reduced parameter count. Without these controls, the numbers in Table I, even if accurately computed, do not establish the paper's central claim.","section":"IV-A and Table I"},{"comment":"The paper repeatedly claims 'computational efficiency' and 'maintaining computational efficiency' (Abstract and Section IV-A.5), but it provides no runtime measurements, no wall-clock time, no complexity analysis of the quantum kernel evaluation, and no comparison with a variational-quantum-circuit LSTM baseline. Section V discusses block-encoding and GPU-based tensor-network simulation as theoretical possibilities, but these are not connected to the concrete implementation used for Table I, and no resource counts are given. The abstract's efficiency claim is therefore unsupported by the experiments.","section":"Abstract, Section V, and Table I text"}],"minor_comments":[{"comment":"'Long short-memory' should be 'Long Short-Term Memory' throughout, including the abstract and the opening sentence of Section I.","section":"Abstract and Section I"},{"comment":"Equation (2) has a garbled display of the Z-score formula; the fraction and the threshold condition should be typeset clearly so that |(x_i - mu)/sigma| > gamma is directly readable.","section":"Eq. (2)"},{"comment":"Figure 2 caption uses a parameter w in U(x_t, w), while the main text defines the feature map as U(v) with no such parameter; the notation should be unified.","section":"Fig. 2"},{"comment":"The hyperparameter table does not list N (the number of reference vectors), the output-layer size, or the quantum encoding details, making the reported parameter total uninterpretable.","section":"Table II and Section IV-B"},{"comment":"The MAPE definition divides by y_i; the paper should state how zero or near-zero AQI values are handled in this metric.","section":"IV-A.3, Eq. (22)"}],"recommendation":"reject","confidential_remarks":"The manuscript is essentially an application note built on the authors' own prior work [23], and it does not include code or data. The central architectural equations are internally inconsistent, the parameter count cannot be reproduced, and the single-run experimental comparison lacks the controls needed to support the claimed quantum advantage. These are load-bearing problems that cannot be fixed by local revisions; the submission would need a substantially reworked model definition and a new experimental methodology. I therefore recommend rejection, despite the potentially interesting direction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe short version: this paper applies the authors' earlier QK-LSTM architecture to a new dataset, but the architecture as written cannot be implemented. The central comparison in Table I is unsupported because the forward pass is undefined.\n\nWhat's new: the AQI benchmark and the performance numbers are new relative to the authors' prior QK-LSTM preprint [23]. The motivation for replacing linear LSTM gates with quantum kernel evaluations is reasonable, and the preprocessing steps (outlier removal, linear interpolation) are standard and clearly described.\n\nWhat's wrong: the gate equations (6a)-(6c) and (6e) produce scalars, but (6d) requires a vector cell state of dimension 16. Nothing lifts the scalar kernel expansions to vector gates. The parameter count is also inconsistent: four gates, each with N kernel weights and a bias, give 4N+4 trainable parameters, which cannot equal 209 for integer N. The quantum circuit in (9)-(12) has no trainable parameters, so the parameter-shift rule in (15) has nothing to act on. And v_t has dimension n+m = 27, while the 4-qubit circuit encodes only 8 angles. These are not minor omissions; the reported model cannot be trained or evaluated as stated.\n\nThe empirical comparison also has weaknesses: a single run with no error bars, no train/test split description, no classical kernel baseline, and efficiency claims without timing measurements. The novelty is incremental because the architecture comes from the authors' own prior work.\n\nThe paper does engage with the literature and the motivation is legitimate, but the internal inconsistency is load-bearing. I don't think this deserves referee time in its current form; it should go back to the authors with a request to define the model properly and rerun the experiments.\n\nIf you want a cautionary example of why equations matter in quantum machine learning, this is a useful case. I wouldn't cite it in my own work.\n\nBest,\n[You]","headline":"A new AQI benchmark, but the QK-LSTM gate equations are internally inconsistent so the reported results are unsupported.","tokens_in":11892,"tokens_out":3071,"would_cite":false,"duration_ms":29857,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Replacing the linear transformations in LSTM cells with quantum kernel evaluations improves AQI forecasting accuracy and cuts trainable parameters from 1,873 to 209.","keywords":["quantum machine learning","quantum kernel methods","long short-term memory","time-series forecasting","air quality index","model compression","NISQ","hybrid quantum-classical computing"],"falsifier":"Run the QK-LSTM on the same Bengaluru dataset with the missing vector construction specified explicitly, and compare the reported RMSE 9.20 against a classical LSTM trained with 209 parameters under identical conditions; if the architecture cannot be instantiated or the accuracy gap disappears, the central claim fails.","tokens_in":10702,"feed_emoji":"🌫️","tokens_out":9983,"duration_ms":90266,"temperature":0.7,"pith_summary":"The paper proposes QK-LSTM, an LSTM variant in which the linear transformations inside the forget, input, candidate, and output gates are replaced by weighted sums of quantum kernel evaluations. On an Air Quality Index forecasting task for Bengaluru, it reports lower error on every metric than a classical LSTM: RMSE 9.20 versus 15.94, MAE 7.15 versus 11.07, MAPE 9.14% versus 13.32%, and $R^2$ 0.84 versus 0.78. It also reports a large parameter reduction, 209 trainable parameters versus 1,873, which it attributes to the expressive power of quantum feature spaces. The motivation is that quantum kernels can capture nonlinear dependencies with fewer parameters than classical recurrent networks, making the model attractive for resource-constrained and near-term quantum settings.","feed_headline":"Quantum kernel LSTM cuts AQI forecast error by 42%","feed_subtitle":"On Bengaluru AQI data, QK-LSTM reaches RMSE 9.20 vs 15.94 with 209 trainable parameters instead of 1,873.","key_machinery":"The load-bearing object is the quantum kernel function $k(v_t, v_j)$, defined as the squared overlap of two data-encoded quantum states. A parameterized circuit $U(v)$ with Hadamard layers, $R_y/R_z$ rotation encoding, and nearest-neighbor CNOT entangling gates maps each concatenated input $v_t=[h_{t-1};x_t]$ to a state $|\\phi(v_t)\\rangle$; the kernel is the measured probability of returning to $|0\\rangle^{\\otimes n}$ after applying $U^\\dagger(v_j)U(v_t)$. This single scalar similarity value replaces the matrix-vector products in each LSTM gate, and the trainable weights $\\alpha_j$ combine the kernel values against a set of reference vectors. The efficiency the paper reports comes from this replacement: the quantum kernel provides the nonlinear feature map, so far fewer classical parameters are needed.","core_discovery":"The central claim is that quantum kernel methods can substitute for learned linear maps in an LSTM cell without losing temporal modeling capacity, and that on the tested AQI dataset this substitution improves accuracy while shrinking the model. The paper defines a QK-LSTM cell where the concatenated input $v_t = [h_{t-1}; x_t]$ is compared against a set of reference vectors through quantum kernels $k(v_t, v_j) = |\\langle \\phi(v_t)|\\phi(v_j)\\rangle|^2$, and each gate becomes $\\sigma(\\sum_j \\alpha_j k(v_t, v_j) + b)$ (or tanh for the candidate). Training uses backpropagation through time for the classical weights and the parameter-shift rule for circuit parameters. The empirical section reports that QK-LSTM outperforms LSTM on RMSE, MAE, MAPE, and $R^2$ while using about one-ninth the parameters, and argues this makes the architecture suitable for NISQ hardware and for simulation on GPUs.","pith_inferences":["A natural next test is to replace the quantum kernel with a classical kernel of the same parameter count; if a classical kernel matches the reported RMSE, the advantage is kernel-based rather than quantum-specific.","The scalar-to-vector gap in the gate equations suggests a concrete repair: define each hidden unit's gate value by giving that unit its own set of reference vectors, or broadcast the scalar kernel sum across the hidden dimension.","If the parameter reduction generalizes, the design principle extends beyond LSTMs: any learned matrix in a recurrent or feedforward cell could be replaced by fixed similarity evaluations against a small reference set, with only combination weights trained.","The paper's NISQ framing implies a cheap benchmark: a four-qubit kernel is small enough to simulate classically, so a GPU tensor-network reproduction of the reported AQI results would test the quantum-hardware claim before any device run."],"forward_implications":["If the reported results hold, QK-LSTM offers a concrete recipe for compressing recurrent models: replace linear gate transformations with kernel expansions, trading parameter count for kernel evaluations.","The architecture can be run partially on classical hardware by simulating the quantum kernel on GPUs, which eases deployment on near-term quantum devices.","The same cell design could be applied to other climate time-series tasks, such as temperature, precipitation, flood, and greenhouse-gas forecasting, since the method is not specific to AQI.","Because the quantum kernel is computed from a 4-qubit circuit, the approach is designed to stay within NISQ resource limits, unlike deeper variational circuits.","Fewer trainable parameters reduce overfitting risk and training cost, which matters for resource-constrained environmental monitoring deployments."],"supporting_citations":[{"why":"Defines the LSTM cell equations that QK-LSTM modifies.","marker":"[5]"},{"why":"Introduces the QK-LSTM architecture that this paper applies to AQI forecasting.","marker":"[23]"},{"why":"Presents the VQC-based quantum LSTM baseline that motivates the kernel-based alternative.","marker":"[14]"},{"why":"Supplies the parameter-shift rule used to train quantum circuit parameters.","marker":"[39]"},{"why":"Provides the feature-Hilbert-space theory behind quantum kernel expressiveness.","marker":"[40]"},{"why":"Is the Bengaluru AQI dataset on which the experiments are run.","marker":"[31]"},{"why":"Defines the NISQ hardware constraints the model is designed to accommodate.","marker":"[17]"}],"fun_headline_variants":["Quantum kernel LSTM beats classical on AQI with 9x fewer parameters","QK-LSTM: Hybrid quantum kernels improve LSTM climate forecasting","Quantum-enhanced LSTM cuts AQI forecast error, shrinks model","Quantum kernel LSTM: Efficient NISQ-ready climate time-series forecasting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the QK-LSTM cell is a well-defined trainable architecture; the paper's gate equations produce scalar values where the cell update needs vectors, and it never states how this is resolved or how many reference vectors are used.","fun_headline_variants_meta":{"raw":{"variants":["Quantum kernel LSTM beats classical on AQI with 9x fewer parameters","QK-LSTM: Hybrid quantum kernels improve LSTM climate forecasting","Quantum-enhanced LSTM cuts AQI forecast error, shrinks model","Quantum kernel LSTM: Efficient NISQ-ready climate time-series forecasting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1364,"prompt_tokens":933,"completion_tokens":431,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":365}},"tokens_in":549,"tokens_out":431,"duration_ms":5240,"temperature":1.0,"reasoning_tokens":365,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:28:39.349637+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the QK-LSTM on the same Bengaluru dataset with the missing vector construction specified explicitly, and compare the reported RMSE 9.20 against a classical LSTM trained with 209 parameters under identical conditions; if the architecture cannot be instantiated or the accuracy gap disappears, the central claim fails.","supporting_citations":[{"cited_title":"Quantum Kernel-Based Long Short-term Memory","cited_arxiv_id":"2411.13225","evidence_quote":"Introduces the QK-LSTM architecture that this paper applies to AQI forecasting."},{"cited_title":"Air quality index (aqi) data","cited_arxiv_id":null,"evidence_quote":"Is the Bengaluru AQI dataset on which the experiments are run."}],"review_version":1}