{"id":"708efe8f-4e51-40e9-9d31-e7d0a3f190bc","arxiv_id":"2411.13225","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"QK-LSTM replaces LSTM gate linear transforms with quantum kernel sums, but its claimed accuracy parity rests on a two-sentence training-only example.","lead":"This paper proposes QK-LSTM, an LSTM variant that replaces linear gate transformations with sums of quantum kernel functions, and tests it on a two-sentence POS tagging example. The authors claim comparable accuracy to classical LSTM with fewer parameters, but the evaluation is too small to support the claim.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Parity claim rests on training accuracy for two hand-picked sentences; no held-out evaluation is reported, so QK-LSTM's LSTM-comparable generalization is unsubstantiated.","rationale":"The reader's weakest assumption is exactly the load-bearing concern: the evaluation uses two handpicked sentences and reports only training curves, so the claim of LSTM-comparable performance is unsupported. I agree with this assessment. The paper explicitly states in Section III-A that two sentences are selected 'for illustrative purposes,' and Section III-B reports performance benchmarking only through Figure 3, which is described as training accuracy and loss. No test accuracy is given anywhere in the manuscript. Since the abstract and discussion generalize from these training curves to statements about model efficiency, compactness, and suitability for deployment, the central claim is not merely under-emphasized but empirically ungrounded. The parameter-shift inconsistency in Section II-B.4 is a real technical flaw—the quantum kernel as defined has no trainable circuit parameters—but it is secondary to the missing generalization evidence. A model with an internally inconsistent training rule could still, in principle, be trained by a different rule; what cannot be rescued is the claim that the reported results demonstrate parity. The proposed concrete test would directly settle whether QK-LSTM generalizes at all on a held-out sentence, and if it fails, the central claim collapses. Therefore the reader's REJECT verdict should stand unchanged.","tokens_in":7569,"tokens_out":2376,"duration_ms":30357,"concrete_test":"Reimplement QK-LSTM and classical LSTM as described and evaluate on a standard POS benchmark with a fixed train/test split, for example the Brown corpus or a Universal Dependencies English treebank. Run at least five random seeds and report held-out token-level accuracy with standard deviation for both models. As a minimal decisive check, train only on 'The dog eat the ice' and test on 'Everybody read that book'; if QK-LSTM does not outperform random classification on the unseen sentence while classical LSTM does, the claimed parity fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the abstract is that QK-LSTM 'achieves performance on par with classical LSTM models, yet with fewer parameters.' The only empirical support is Section III-A/III-B and Figure 3, which report training accuracy and loss for two manually selected sentences: 'The dog eat the ice' and 'Everybody read that book.' There is no train/test split, no held-out sentence, no repeated initialization, and no error bars. Training curves on two sequences can look identical even when one model memorizes and the other generalizes, so the claimed parity is not established. A second, independent issue is that the parameter-shift rule in Eqs. (9)-(10) has no trainable quantum-circuit parameter to shift: the feature map U(v) in Eq. (7) encodes input data through rotations whose angles are functions of v, but no parameter in U(v) is learned. Thus the stated quantum-circuit gradient is not well-defined as written. The compression comparison is also fragile because the architecture replaces full gate matrices with a fixed number of kernel weights, so the parameter count depends on the number of reference vectors N, which is never specified. The decisive weakness, however, is the absence of any generalization test.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Quantum Kernel-Based Long Short-Term Memory (QK-LSTM), which replaces the linear transformations in each LSTM gate with weighted sums of quantum kernel functions evaluated between the concatenated input and a set of reference vectors. The authors report an application to Part-of-Speech tagging on two manually selected sentences, showing training accuracy and loss curves that they interpret as achieving performance on par with classical LSTM and QLSTM while using fewer trainable parameters (183 versus 477). The manuscript also discusses parameter-shift-rule gradients for the quantum circuit and positions the architecture for NISQ/edge deployment.","tokens_in":7807,"tokens_out":7436,"duration_ms":73187,"significance":"The conceptual idea of substituting learned matrix multiplications with fixed quantum kernel expansions is worth exploring, and the paper identifies a relevant application area in sequence modeling. The explicit parameter-count comparison is a useful starting point, but the current manuscript does not provide a trustworthy evaluation: the only empirical evidence is in-sample fitting on two sentences, and the model equations contain an unresolved dimensional inconsistency. The contribution is therefore preliminary. No code, data, or machine-checked proofs are provided, so the claims cannot be independently verified.","major_comments":[{"comment":"The central claim that QK-LSTM 'achieves performance on par with classical LSTM models' is supported only by training accuracy and loss curves on two hand-picked sentences ('The dog eat the ice' and 'Everybody read that book'). There is no held-out test set, no repeated initialization, and no error bars. Reporting the training loss after fitting the model on the same two sentences demonstrates curve fitting, not generalization, so the parity claim in the abstract and Section III-B is not established. A standard POS benchmark with a train/test split and multiple seeded runs is required.","section":"III-A, III-B, Fig. 3"},{"comment":"The gate equations are dimensionally inconsistent. Since k(vt,vj) in Eq. (8) is a scalar and the sums in Eqs. (3a)-(3e) are over scalar weights times scalar kernels, each gate activation is a scalar, not a vector. However, an LSTM hidden state ht and cell state Ct must be vectors, and Table I specifies a hidden dimension of 6. As written, Eq. (3f) produces a scalar hidden state, so the architecture cannot represent the multi-dimensional state that the LSTM framework requires. The authors need to define vector-valued kernels, per-output-dimension weights, or some other mechanism that preserves the hidden dimension.","section":"II-B-2, Eqs. (3a)-(3f), Table I"},{"comment":"The parameter-shift rule is stated for 'a circuit parameter θ', but no trainable circuit parameter is defined in the feature map. In Eq. (7), Uenc(v) contains rotation angles that are functions of the input v and Uent is a fixed CNOT entangler; neither contains a learned parameter. Consequently, Eq. (13) updates an undefined θ, and the gradient formula in Eqs. (9)-(10) has no object to act on. If the quantum kernel is meant to be fixed after data encoding, the parameter-shift discussion should be removed and training should be over α and b only; if trainable circuit parameters are intended, they must be introduced explicitly in U(v) and the parameter-shift derivation must be repeated for that specific parameterization.","section":"II-B-4-c, Eqs. (9)-(10)"},{"comment":"The compression claim is not reproducible because N, the number of reference vectors, is never specified. The trainable parameters in the QK-LSTM include the weights α_j^(gate) for each of the four gates, and the parameter count therefore scales with N; without N and an explicit counting formula, the reported value of 183 trainable parameters cannot be verified. The authors should state N and provide the exact parameter-count expressions for QK-LSTM and LSTM, including how biases and output layers are counted.","section":"II-B-2, Table I"}],"minor_comments":[{"comment":"POS labels are given for 'The dog eat the ice' but not for the second sentence 'Everybody read that book'; please provide the full labeling or clarify that the second sentence is used without explicit labels.","section":"III-A"},{"comment":"The caption introduces U(x_i, w) and a parameter w that are not defined in the text; the notation should be consistent with Eq. (7).","section":"Fig. 2(b)"},{"comment":"The heading 'Performance Benchmmarking' contains a typo; it should be 'Performance Benchmarking'.","section":"III-B"},{"comment":"The four kernel functions k(f), k(i), k(C), and k(o) are called gate-specific, but Eq. (7) defines a single feature map U(v) for all gates; please explain how the four kernels differ or use one common kernel with gate-specific weights.","section":"Eq. (3)"},{"comment":"Reference [40] cites the PennyLane documentation; the parameter-shift rule should be attributed to the original works (e.g., Mitarai et al. or Schuld et al.) if the rule is retained.","section":"References"},{"comment":"Phrases such as 'robust loss minimization' and 'efficient convergence' are overstatements when based on a single training run on two sentences; statistical support is needed.","section":"Abstract and III-B"}],"recommendation":"reject","confidential_remarks":"The manuscript is at an early stage: the model definition is incomplete and the empirical section is anecdotal. Even though some issues (e.g., the missing test set) could be addressed in principle, the combination of the undefined trainable circuit parameters, the scalar-versus-vector inconsistency, and the absence of any generalization experiment makes the current submission unsuitable for publication. A substantially rewritten version with a well-defined vectorized architecture and a real benchmark would be needed for reconsideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: this paper has a new architecture—quantum kernels used as weighted sums inside LSTM gates—but its central claim of LSTM-parity is built on training accuracy for two sentences, with no held-out test. The parity claim doesn't hold as stated.\n\nWhat's genuinely new: to my knowledge nobody has replaced the linear transformations in each LSTM gate with weighted sums of quantum kernel evaluations. The exposition is clean and the idea is coherent: you trade dense weight matrices for a fixed set of kernel weights, which could give compression if the kernel is expressive enough. The authors also engage the relevant prior work on QLSTM and quantum kernels, and they correctly frame the contribution as an incremental architecture rather than a paradigm shift.\n\nWhere it falls down: the evaluation. Section III uses two hand-picked sentences ('The dog eat the ice' and 'Everybody read that book'), reports training accuracy and loss over 100 epochs, and calls that a benchmark. There is no train/test split, no repeated initialization, no error bars. Training curves can match even when one model memorizes and the other generalizes, so 'performance on par' is not established. This is the load-bearing flaw, and the reader's stress-test is right.\n\nThere's also a technical sloppiness in the parameter-shift subsection. Equations (9)–(10) claim to take the gradient of the kernel with respect to a circuit parameter θ, but no trainable circuit parameter is defined: the feature map U(v) has rotation angles that are functions of the input v, not learned variables. That section is not actually used in the experiments, which only train α and b, so it's a minor inconsistency rather than a fatal one, but it should be fixed.\n\nThe compression comparison is also partly a consequence of the chosen tiny dimensions (embedding 8, hidden 6), and the number of reference vectors N is never specified, so the parameter count comparison is not conclusive.\n\nBottom line: the architecture is a reasonable thing to explore, and the paper is clearly written, but the empirical evidence as it stands doesn't support the headline claim. A serious referee could send this back for major revision: require a real dataset, held-out evaluation, multiple seeds, and a clarification of the gradient story. I'd bring it to a reading group as an example of how easy it is to over-read training curves in QML papers, but I wouldn't cite it yet.\n\nRecommendation: send to review if the venue wants to flag fixable empirical weaknesses; otherwise desk reject for insufficient evidence.","headline":"New QK-LSTM architecture, but the parity claim rests on training curves for two sentences; the idea is worth exploring, the evidence isn't.","tokens_in":8304,"tokens_out":2983,"would_cite":false,"duration_ms":31142,"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":"Replacing an LSTM's linear gate operations with quantum kernel evaluations produces a smaller model that matches the classical network's training accuracy on a part-of-speech tagging task.","keywords":["quantum kernel methods","long short-term memory","sequence modeling","model compression","part-of-speech tagging","quantum machine learning","NISQ-era devices"],"falsifier":"Train QK-LSTM and a classical LSTM on a standard part-of-speech corpus with a held-out test split and compare test accuracy at matched parameter counts; if QK-LSTM falls materially below LSTM, the on-par claim is refuted.","tokens_in":7358,"feed_emoji":"⚛️","tokens_out":9130,"duration_ms":82475,"temperature":0.7,"pith_summary":"This paper sets out to show that quantum kernel functions can stand in for the linear transformations inside an LSTM cell without sacrificing sequence-modeling performance. It introduces QK-LSTM, in which each gate's usual weighted sum is replaced by a weighted sum of kernel similarities between the current input-hidden concatenation and a set of reference vectors. On a part-of-speech tagging exercise, the paper reports that QK-LSTM reaches training accuracy and loss curves comparable to classical LSTM and to a variational-circuit QLSTM while using 183 trainable parameters versus 477. The point of the exercise is model compression: if quantum feature spaces can carry the same expressive load with fewer parameters, sequence models become more feasible on NISQ-era and edge devices.","feed_headline":"Quantum kernel LSTM matches LSTM accuracy with fewer parameters","feed_subtitle":"Replacing LSTM's linear gates with quantum kernel functions shrinks trainable parameters from 477 to 183 on a POS-tagging task.","key_machinery":"The central object is the quantum kernel function $k(v_t,v_j)=|\\langle\\phi(v_t)|\\phi(v_j)\\rangle|^2$, a similarity measure in the quantum feature space induced by the circuit $U(v)$. It carries the argument by replacing the linear transformations in each LSTM gate with a kernel-weighted sum over reference vectors, so the model's non-linearity and expressiveness come from the quantum feature map rather than from a large weight matrix. The parameter-shift rule makes this kernel differentiable, allowing the same backpropagation-through-time training loop as a classical LSTM.","core_discovery":"QK-LSTM replaces the linear map $W[h_{t-1},x_t]+b$ in the forget, input, cell, and output gates of a standard LSTM with a sum $\\sum_j \\alpha_j k(v_t, v_j)+b$, where $v_t=[h_{t-1},x_t]$ and $k(v_t,v_j)=|\\langle\\phi(v_t)|\\phi(v_j)\\rangle|^2$ is a quantum kernel evaluated by a small parameterized circuit with Hadamard initialization, $R_y/R_z$ data encoding, and CNOT entanglement. The trainable weights $\\alpha_j$ and biases are updated by backpropagation through time, and circuit-level gradients use the parameter-shift rule. On a two-sentence part-of-speech tagging benchmark, the model reaches training accuracy and loss curves comparable to classical LSTM and QLSTM, with 183 trainable parameters against the classical LSTM's 477. The paper's central claim is that quantum kernel evaluations are expressive enough to replace the linear gate transformations, compressing the model while keeping its learning dynamics intact.","pith_inferences":["The reported parity is measured on training curves for two hand-picked sentences, so the paper does not yet establish generalization; a held-out evaluation on a larger POS-tagged corpus would test whether the quantum kernel's inductive bias helps or hurts.","If the circuit parameters are also trained (the paper lists them as parameters), the true parameter count and quantum-resource cost could exceed 183, so the compression ratio depends on how the kernel circuit parameters are counted.","The reference vectors $v_j$ can be chosen from training data or learned; a natural test is to vary $N$ and measure how the accuracy-parameter trade-off changes, which would show how much of the compression is due to the quantum feature map versus the choice of reference points.","A classical kernel LSTM with the same kernel-weighted gate structure would isolate the quantum contribution; without that baseline, the improvement cannot be attributed to quantum resources alone."],"forward_implications":["Sequence models can be compressed to roughly a third of the trainable parameters while preserving training performance on the reported task.","The standard LSTM training pipeline remains intact, so QK-LSTM can reuse backpropagation-through-time and parameter-shift updates without new optimization machinery.","Shallow kernel circuits, rather than deep variational ansatze, may be sufficient for quantum-enhanced sequence modeling, lowering the hardware requirement.","The approach extends in principle to other sequential tasks named in the paper, including time-series forecasting and signal classification."],"supporting_citations":[{"why":"Defines the LSTM cell equations that QK-LSTM modifies by replacing linear gate transformations with quantum kernels.","marker":"[2]"},{"why":"Supplies the VQC-based QLSTM baseline that the paper compares against in the accuracy and loss curves.","marker":"[23]"},{"why":"Provides the parameter-shift rule used to compute gradients through the quantum kernel circuits during training.","marker":"[40]"},{"why":"Introduces the tailored quantum kernel construction that the gate computations are based on.","marker":"[26]"},{"why":"Establishes the inner-product form of quantum kernels used in the kernel evaluation equation.","marker":"[27]"},{"why":"Establishes the quantum NLP context and POS-tagging style task used as the benchmark.","marker":"[20]"}],"fun_headline_variants":["Quantum kernel LSTM cuts parameters 2.6x with same accuracy","Quantum kernel LSTM: 62% fewer parameters, same accuracy","Quantum kernels compress LSTM to 183 params on POS tags","QK-LSTM: quantum kernel gates match LSTM with 183 weights","Quantum kernel LSTM: fewer params, same sequence performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The parity claim leans on training curves from two manually chosen sentences with no held-out test set, so the reported accuracy could reflect fitting those particular sentences rather than general sequence-modeling ability.","fun_headline_variants_meta":{"raw":{"variants":["Quantum kernel LSTM cuts parameters 2.6x with same accuracy","Quantum kernel LSTM: 62% fewer parameters, same accuracy","Quantum kernels compress LSTM to 183 params on POS tags","QK-LSTM: quantum kernel gates match LSTM with 183 weights","Quantum kernel LSTM: fewer params, same sequence performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000565,"raw_usage":{"total_tokens":2672,"prompt_tokens":933,"completion_tokens":1739,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":1648}},"tokens_in":549,"tokens_out":1739,"duration_ms":11863,"temperature":1.0,"reasoning_tokens":1648,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:40:00.657460+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train QK-LSTM and a classical LSTM on a standard part-of-speech corpus with a held-out test split and compare test accuracy at matched parameter counts; if QK-LSTM falls materially below LSTM, the on-par claim is refuted.","supporting_citations":[],"review_version":1}