{"id":"0bc68317-a3b5-419d-af82-e62811037baa","arxiv_id":"2507.08255","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Swapping a classical embedding for an IQP quantum circuit embedding improved LLM-based table imputation in simulations, but the gains are reported without code, data, or error bars.","lead":"This paper proposes Quantum-UnIMP, a system that uses a small quantum circuit to generate input representations for a large language model that fills in missing table values. On three datasets it reports better imputation accuracy than the classical version, but the work provides no code, data, or uncertainty estimates.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed quantum advantage is untestable: the 8-qubit IQP embedding compresses d features to 8 Pauli-Z expectations via an unspecified mapping, and the ablation lacks a matched classical nonlinear control.","rationale":"The reader's weakest-assumption identification is essentially correct: the framework assumes that a fixed 8-qubit IQP circuit read out through Pauli-Z expectations provides a strictly more expressive embedding than an MLP. My stress-test locates the precise failure mode in Sections 3.2.2 and 3.2.4: the architecture compresses all inputs to eight real numbers and never specifies how d > 8 features are mapped onto the circuit parameters. Because the readout consists only of single-qubit Z expectations, any entanglement created inside the circuit is invisible to the downstream LLM; the output is a classically computable 8-dimensional nonlinear function of the input, not a certificate of quantum expressivity. The ablation in Section 5.2 does not include a matched classical nonlinear embedding with the same output dimension, so the comparison cannot isolate the quantum component. In addition, the absence of code, data, seeds, and error bars makes the numerical claims unverifiable. I agree with the reader's REJECT verdict, and my concern reinforces it rather than changing it; hence UNCHANGED. I do not raise objections about the authenticity of the authors; the issue is purely that the central causal claim is not supported by the evidence presented.","tokens_in":9827,"tokens_out":5617,"duration_ms":67373,"concrete_test":"Re-implement the pipeline with a fully specified compression from each dataset's d features to the 8 qubit parameters, then train Quantum-UnIMP and a matched classical control in which the IQP circuit is replaced by a fixed nonlinear map with the same 8-dimensional output (e.g., 8 random Fourier features computed from the same preprocessed d features), holding all downstream hyperparameters, missingness masks, and seeds fixed. If the classical control reproduces Quantum-UnIMP's gains within error bars, the attribution of the improvement to the quantum feature map is refuted; if not, the concern is mitigated, though published code/data with multiple seeds would still be needed to verify the reported numbers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the fixed IQP feature map is the key driver of the reported gains—breaks at the compression/readout step. Section 3.2.2 states that a d-dimensional feature vector is linearly mapped to the parameters of an 8-qubit circuit, but no mapping is specified; Adult has 14 features, Bank 21, and Synthetic 20 plus 384 text-embedding dimensions, all with d > 8. The map must therefore discard or mix features before any quantum processing. Section 3.2.4 then forms the embedding as n = 8 Pauli-Z expectations, so the LLM receives an 8-dimensional vector regardless of the input. Single-qubit Z expectations are insensitive to entanglement and to correlations between qubits, so the claimed mechanism of superposition and entanglement cannot be responsible for any improvement that survives this readout. The ablation in Section 5.2 compares only against a random projection and the classical MLP; it does not control for the change in output dimensionality, nor for a fixed classical nonlinear embedding of the same 8 dimensions. The reported 15.2% RMSE and 8.7% F1 gains could therefore be an artifact of compression or of a nonlinear random-feature map rather than of quantum expressivity. The claim is also unverifiable as written: no code, data, seeds, or error bars are provided, and the abstract percentages do not match Table 1 exactly.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Quantum-UnIMP, a hybrid quantum-classical imputation framework that replaces the classical input embedding of the UnIMP model with a fixed 8-qubit IQP feature map. The classical feature vector is preprocessed, mapped to gate parameters, and the Pauli-Z expectation values of the measured state are fed as token embeddings into a hypergraph transformer. Experiments on Adult, Bank, and a synthetic healthcare dataset report lower RMSE and higher F1 than MICE, MissForest, GAIN, and UnIMP, with an ablation on Adult claiming the quantum embedding is the key driver. The paper also includes t-SNE visualizations and a discussion of NISQ limitations.","tokens_in":10068,"tokens_out":7388,"duration_ms":76933,"significance":"The idea of using IQP feature maps as a drop-in replacement for classical embeddings in LLM-based tabular imputation is potentially interesting, and the random-projection ablation is a sensible first step. However, the paper as written does not establish the central claim: the feature-to-qubit compression is unspecified, the Pauli-Z readout is insensitive to the correlations the paper invokes, the ablation lacks a matched classical nonlinear control, and the quantitative results are not reproducible because no code, data, seeds, or error bars are provided. The abstract's headline percentages are also inconsistent with Table 1. If the results hold with proper controls, they would be a valuable empirical contribution to hybrid quantum-classical ML.","major_comments":[{"comment":"The dimensionality reduction from the classical feature vector to the 8-qubit circuit is never specified. All three datasets have d > 8 (Adult has 14 features, Bank 21, and Synthetic has 20 features plus 384-dimensional text embeddings from all-MiniLM-L6-v2), and the display in §3.2.4 produces an 8-dimensional embedding xq. The paper states only that 'a linear mapping is established' from xc to the gate parameters θ, without defining the mapping or stating whether it is trained. Because the LLM receives only xq, the reported gains could be an artifact of aggressive dimensionality reduction rather than of the quantum circuit processing the full input. This is load-bearing for the claim that the IQP map is the key driver.","section":"§3.2.2–3.2.4"},{"comment":"The embedding xq consists exclusively of single-qubit Pauli-Z expectation values. These are single-qubit marginals and are not directly sensitive to joint correlations between qubits; they therefore cannot by themselves represent the entanglement or higher-order correlations that Sections 5.2 and 6.1 use to explain the gains. To substantiate the claimed mechanism, the measurement should include multi-qubit operators (e.g., ZZ or ZZZ terms), or the interpretation should be revised to describe xq as a classical nonlinear feature map rather than as an entanglement-based representation.","section":"§3.2.4 and §6.1"},{"comment":"The ablation does not control for the change in output dimensionality or for the nonlinearity of the embedding. UnIMP (Random Proj.), UnIMP (Classical MLP), and Quantum-UnIMP differ in both the type of map and the dimensionality of the representation fed to the LLM. A matched control using a fixed classical nonlinear feature map of the same 8-dimensional output (e.g., random Fourier features or a fixed random MLP layer) is needed to attribute the observed improvement to the quantum origin of the features. The statement that the results 'unequivocally prove that the quantum feature map is the key driver' is therefore not supported by the evidence in Table 2.","section":"§5.2, Table 2"},{"comment":"The quantitative claims in the abstract do not match the reported results. From Table 1, the RMSE reductions of Quantum-UnIMP relative to UnIMP are 13.8% (Adult), 14.7% (Bank), and 15.6% (Synthetic), and the F1 improvements are 7.5%, 9.5%, and 11.3%; the abstract's '15.2%' and '8.7%' do not correspond to any of these values. Moreover, no standard deviations, number of runs, or seeds are reported, and no code or data are provided, so the statistical significance and reproducibility of the central empirical claim cannot be assessed.","section":"Abstract and Table 1"}],"minor_comments":[{"comment":"Section headings contain duplicated numbers (e.g., '2.1 3.1.', '3.1 4.1.', '4.1 5.1.', '5.1 6.1.'), which appear to be a numbering error and should be corrected.","section":"Throughout"},{"comment":"'an ubiquitous problem' should be 'a ubiquitous problem'.","section":"§1.1"},{"comment":"Minor spacing errors appear in 'Numerical F eatures', 'Categorical F eatures', and 'T ext F eatures'.","section":"§3.2.1"},{"comment":"Figure 3 is referenced for the t-SNE visualization, but no figure is included in the manuscript text; either include the figure or remove the reference.","section":"§5.3"},{"comment":"The baselines are described as 'state-of-the-art' but only one LLM-based method (UnIMP) is included; other recent LLM-based tabular imputation methods are not cited or compared.","section":"§4.2"}],"recommendation":"reject","confidential_remarks":"The manuscript is not in a reproducible state, and the central mechanism claim is unsupported by the described architecture. I would be willing to reconsider a substantially revised version that specifies the compression, adds a matched classical nonlinear control, reports error bars and reproducibility artifacts, and corrects the internal numerical inconsistencies."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a direct substitution of an IQP feature map for the MLP embedding inside UnIMP, and the paper's central claim—that the quantum embedding drives the gains—is not supportable from what's in the manuscript. The specific combination is new, and the writing is mostly clear, but the evidence is missing a load-bearing part.\n\nWhat's new: the idea of using an IQP circuit as the input embedding for an LLM-based imputer is a reasonable thing to try, and I don't see it in the cited literature. The experimental setup is straightforward: three mixed-type datasets, standard baselines, and the results are presented cleanly. The limitations section honestly acknowledges simulator cost and NISQ noise.\n\nWhere it falls apart: First, the mechanism. The paper attributes the improvement to superposition and entanglement, but the embedding is the vector of single-qubit Pauli-Z expectations. Those are local observables; they don't see entanglement between qubits. So whatever gain is claimed cannot be explained by the quantum correlations the text leans on. Second, the compression. Adult has 14 features, Bank 21, Synthetic 20 plus text embedding dimensions, yet the circuit is 8 qubits and the output is 8 numbers. Section 3.2.2 says there is a linear mapping from x_c to gate parameters, but never specifies it. That is a big missing piece. Third, the ablation. It compares against a random projection and the classical MLP, but not against a fixed classical nonlinear embedding of the same 8 dimensions. The gains could come from the change in output dimensionality or from any nonlinear random feature map, not from anything quantum. Fourth, the abstract percentages don't match Table 1: the table shows a max RMSE reduction of 15.6% on Synthetic, not 15.2%; the max F1 improvement is 11.3%, not 8.7%. Finally, no code, data, seeds, or error bars are provided, so nothing is independently checkable.\n\nWho it's for: someone surveying attempts to graft quantum embeddings onto classical pipelines might skim it, but no one should build on the current numbers.\n\nRecommendation: desk-reject this version. The idea is worth an experiment, but it needs a fully specified encoding, a matched classical nonlinear control, and complete reproducibility before it deserves referee time.","headline":"A plausible-sounding but unverifiable embedding swap; the quantum mechanism is undercut by the local readout and the empirical claims can't be checked.","tokens_in":10606,"tokens_out":3144,"would_cite":false,"duration_ms":34359,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that replacing classical input embeddings with an 8-qubit IQP quantum feature map reduces LLM-based imputation error by up to 15.2% (RMSE) and improves categorical imputation F1 by up to 8.7% on mixed-type tabular…","keywords":["quantum machine learning","data imputation","large language models","quantum feature maps","IQP circuits","hybrid quantum-classical algorithms","mixed-type data","missing data"],"falsifier":"Re-run the same pipeline with the IQP circuit replaced by a fixed classical nonlinear embedding of the same output size (for example, random Fourier features or a random ReLU projection to eight dimensions), keeping the downstream transformer identical. If the classical replacement reproduces the reported RMSE and F1 gains, the quantum mechanism is not the driver; the paper's argument would also be weakened if the 15.2% gap on the synthetic healthcare set disappears when the text-derived features are removed from the encoding.","tokens_in":9583,"feed_emoji":"⚛️","tokens_out":7559,"duration_ms":74389,"temperature":0.7,"pith_summary":"The paper tries to establish that the input embedding, not the language model, is the bottleneck for LLM-based tabular imputation, and that a shallow quantum feature map removes that bottleneck. It proposes Quantum-UnIMP, which replaces the classical MLP embedding of an existing LLM imputer with an 8-qubit Instantaneous Quantum Polynomial (IQP) circuit; the circuit's Pauli-Z expectation values become the input representations for a hypergraph Transformer that predicts masked cell values. On the Adult Income, Bank Marketing, and a synthetic healthcare dataset with non-random missingness, the framework reports up to 15.2% lower RMSE for numerical imputation and up to 8.7% higher macro F1 for categorical imputation against classical, GAN-based, and LLM-based baselines. An ablation that swaps the quantum map for a random projection supports the claim that the quantum embedding itself drives the improvement.","feed_headline":"Quantum embedding cuts language-model imputation error by 15%","feed_subtitle":"Swapping classical embeddings for a shallow quantum map also lifts categorical F1 by 8.7% on mixed-type tabular benchmarks.","key_machinery":"The central object is the IQP feature map $U(x)=H^{\\otimes n}U_{\\mathrm{diag}}(x)H^{\\otimes n}$, a shallow circuit whose diagonal phase gates are controlled by the input features; the paper reads out the embedding as the vector of Pauli-Z expectation values $x_q=[\\langle Z_1\\rangle,\\ldots,\\langle Z_n\\rangle]$ on $n=8$ qubits. This fixed, non-variational circuit replaces the classical MLP embedding and feeds the Transformer-based hypergraph LLM that performs masked-token imputation. The map's claimed role is to translate mixed-type rows, including text-derived vectors, into a representation where class-relevant structure is more separable, as illustrated by t-SNE plots.","core_discovery":"The central claim is that a fixed, non-trained IQP embedding circuit is a strictly more effective input representation for LLM-based imputation than the classical learned embedding it replaces. The paper reports consistent gains across three mixed-type benchmarks: RMSE drops from 0.29 to 0.25 on Adult, 0.34 to 0.29 on Bank, and 0.45 to 0.38 on the synthetic healthcare set, while macro F1 rises from 0.80 to 0.86, 0.74 to 0.81, and 0.71 to 0.79 respectively. The authors attribute the gains to the IQP map's ability to encode higher-order, entangled correlations that classical linear or MLP embeddings miss, and they take the ablation result (random projection 0.36 RMSE, classical MLP 0.29, quantum 0.25) as evidence that the quantum map is the causal driver.","pith_inferences":["The paper does not compare the IQP embedding against other fixed nonlinear classical embeddings of similar dimension, so whether entanglement specifically is responsible, rather than any high-dimensional nonlinear map, remains an open question.","Since the 8-qubit circuit compresses all input features, the method implicitly tests whether a very low-dimensional summary can preserve the correlations needed for imputation; if true, cheaper classical kernel methods may be able to mimic the result.","A natural extension the author does not run is scaling the number of qubits or adding trainable parameters to the embedding; monotonic gains with qubit count would support the quantum expressivity story, while flat gains would suggest the effect saturates at the representation dimension."],"forward_implications":["If the central claim holds, any LLM-based tabular imputer can be improved by swapping its classical input embedding for a fixed shallow IQP embedding, with no change to the downstream transformer.","The gains are largest on the synthetic dataset designed with non-random missingness, so quantum embeddings would matter most where missingness depends on hidden correlations.","Because the embedding circuit is fixed and only 8 qubits, the approach is compatible with near-term quantum hardware, provided noise can be controlled.","The ablation implies that random or raw projections are far worse, so embedding quality, not model capacity, is what drives imputation accuracy in this architecture."],"supporting_citations":[{"why":"statistical chained-equations baseline whose RMSE/F1 the proposed method must beat.","marker":"[18]"},{"why":"GAN-based imputation baseline included in the comparison tables.","marker":"[19]"},{"why":"the LLM-based imputation architecture whose classical MLP embedding is replaced by the quantum map.","marker":"[20]"},{"why":"introduces the IQP quantum feature-map construction the embedding circuit implements.","marker":"[21]"},{"why":"random-forest imputation baseline compared on all three datasets.","marker":"[22]"},{"why":"supplies the quantum simulator backend used to run the circuit experiments.","marker":"[24]"}],"fun_headline_variants":["Quantum-UnIMP: LLM imputation error down 15.2% with IQP maps","IQP quantum embeddings improve LLM imputation: 15.2% lower RMSE","Quantum LLM imputation: 15.2% RMSE drop, 8.7% F1 gain","Shallow IQP quantum map cuts LLM imputation error 15.2%","Quantum-accelerated LLM imputation: 15% error cut, 8.7% F1 gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a fixed 8-qubit quantum circuit, read out as eight simple measurements, gives the LLM a strictly better description of each row than the ordinary learned embedding it replaces, even though every row's many features—numerical, categorical, and text-derived—must be squeezed into those eight rotation angles.","fun_headline_variants_meta":{"raw":{"variants":["Quantum-UnIMP: LLM imputation error down 15.2% with IQP maps","IQP quantum embeddings improve LLM imputation: 15.2% lower RMSE","Quantum LLM imputation: 15.2% RMSE drop, 8.7% F1 gain","Shallow IQP quantum map cuts LLM imputation error 15.2%","Quantum-accelerated LLM imputation: 15% error cut, 8.7% F1 gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000735,"raw_usage":{"total_tokens":3298,"prompt_tokens":967,"completion_tokens":2331,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":2205}},"tokens_in":583,"tokens_out":2331,"duration_ms":15940,"temperature":1.0,"reasoning_tokens":2205,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:22:41.571431+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same pipeline with the IQP circuit replaced by a fixed classical nonlinear embedding of the same output size (for example, random Fourier features or a random ReLU projection to eight dimensions), keeping the downstream transformer identical. If the classical replacement reproduces the reported RMSE and F1 gains, the quantum mechanism is not the driver; the paper's argument would also be weakened if the 15.2% gap on the synthetic healthcare set disappears when the text-derived features are removed from the encoding.","supporting_citations":[{"cited_title":"Multiple imputation by chained equations: what is it and how does it work?","cited_arxiv_id":null,"evidence_quote":"statistical chained-equations baseline whose RMSE/F1 the proposed method must beat."},{"cited_title":"Gain: Missing data imputation using generative adversarial nets,","cited_arxiv_id":null,"evidence_quote":"GAN-based imputation baseline included in the comparison tables."},{"cited_title":"Supervised learning with quantum-enhanced feature spaces,","cited_arxiv_id":null,"evidence_quote":"introduces the IQP quantum feature-map construction the embedding circuit implements."},{"cited_title":"Missforest–non-parametric missing value imputation for mixed-type data,","cited_arxiv_id":null,"evidence_quote":"random-forest imputation baseline compared on all three datasets."},{"cited_title":"Pennylane: Automatic differen- tiation of hybrid quantum-classical computations,","cited_arxiv_id":null,"evidence_quote":"supplies the quantum simulator backend used to run the circuit experiments."}],"review_version":1}