{"id":"71c108c7-83aa-4e7c-aec7-79b3e4e4b51a","arxiv_id":"2607.26925","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"Voxelized molecular geometries encoded as equal superpositions over sparse computational-basis states recover A atoms with O(A log A) shots via coupon-collector sampling, shown on IBM hardware for ethylamine.","lead":"The paper encodes molecules as sparse equal superpositions over voxel-and-atom-type basis states so geometry can be read out by ordinary computational-basis sampling instead of full tomography. A 10-atom ethylamine demo on IBM Kingston recovers most atoms in a few hundred shots, offering a practical readout format for near-term quantum generative chemistry models.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Oracle recall is shown; unsupervised support recovery under hardware ground-state bias is not, so the generative readout pathway is unproven.","rationale":"The reader correctly flags state preparation as out of scope and the ~1213-gate isometry as non-scalable; that bounds any end-to-end “practical NISQ molecular encoding” claim and already justifies CONDITIONAL. The scoped math claim (coupon collector once an equal-superposition support state exists) is sound, and the hardware oracle-recall numbers are a real, well-documented demo. The additional load-bearing gap is on the decoding side the paper advertises for generative modeling: recovery without knowing M*. R is oracle-only; SNR is mislabeled as ground-truth-free; Fig. 2a’s ground-state bias makes unsupervised top-k recovery a live failure mode that is never tested. That does not overturn the demo, but it tightens the same CONDITIONAL verdict: accept as a readout-encoding observation with an honest prep caveat, not as a validated molecule-from-shots pipeline on NISQ. No change to REJECT/ACCEPT; keep CONDITIONAL with both prep and unsupervised-identification caveats explicit.","tokens_in":18074,"tokens_out":669,"duration_ms":66233,"concrete_test":"Using the S=200 IBM Kingston shot histograms (re-run if raw counts unavailable), rank all 2^8 outcomes by empirical frequency; take the top-10 (and top-15) indices and compute overlap with the 10 true c_a from Table 1. If mean unsupervised recovery is materially below reported R̄≈0.98 (e.g., <0.7), oracle recall overstates usable NISQ reconstruction and the Sec. 5 coupon-collector generative pathway does not hold as stated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest empirical claim is high reconstruction recall (R̄=0.94 at 116 shots, 0.98 at 200) on IBM Kingston. R in Eq. (15) is an oracle metric: it scores only whether known atoms in M* appear at least once and ignores false positives. For the intended use (Sec. 5 generative pipeline; Sec. 4.3), the support is unknown and must be inferred from the histogram. The paper calls SNR (Eq. 16) a “ground-truth-free proxy,” but Eq. 16 is defined with labeled partitions ⟨N_valid⟩ over the 10 true atoms and ⟨N_empty⟩ over the 182 empties—so it measures separability given M*, not recovery without M*. Fig. 2a shows strong hardware bias toward low-index/ground states from T1 leakage. True atom indices (Table 1: 16,25,29,70,78,94,100,159,163,169) are scattered; nothing in the paper shows that frequency ranking (top-A or thresholded) recovers those indices rather than decoherence-favored empty voxels. If it does not, O(A log A) shots yield high oracle recall while still failing to reconstruct an unknown molecule—the load-bearing step for the claimed NISQ generative readout.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes quantum voxelization encoding: molecular geometries are discretized onto a 3D voxel grid, each atom’s voxel index and chemical type are packed into a single computational-basis label c_a, and the molecule is represented as an equal superposition over the A occupied basis states. Once that sparse support state is prepared, reconstruction reduces to classical coupon-collector sampling in the computational basis, requiring O(A log A) shots in the noise-free case (Sec. 3.2, App. C). The authors assume state preparation exists and focus on readout. They encode 10-atom ethylamine into 8 qubits (V=4, T=3, C=192), run on IBM Kingston and a noisy simulator, and report mean reconstruction recall R̄=0.94 at 116 shots and R̄=0.98 at 200 shots despite ~70% wasted shots, arguing this is a practical, readout-efficient representation for NISQ molecular generative models.","tokens_in":18352,"tokens_out":1457,"duration_ms":30568,"significance":"If the encoding is usable end-to-end, converting geometry readout from exponential tomography to polynomial support sampling is a genuine and useful contribution for quantum generative chemistry on NISQ hardware. The coupon-collector analysis is standard and correctly applied; the hardware campaign is concrete (10 repeats per shot budget, wasted-shot and noise breakdowns, simulator baseline, device appendix). Explicit acknowledgment that preparation is out of scope and that spatial precision is lost to the grid is appropriate. The work is best read as a readout-compatible representation and a small-scale NISQ feasibility demo, not as a complete data-loading or generative pipeline.","major_comments":[{"comment":"Sec. 4.3 and Eq. (15): reconstruction recall R is an oracle metric—it scores only whether known atoms in M* appear at least once and does not penalize false positives. The intended generative use case (Sec. 5 pipeline; end of Sec. 4.3) requires inferring an unknown support from the histogram alone. The paper never reports unsupervised recovery (e.g., top-A frequency ranking, thresholding, or clustering) of the true indices in Table 1. With the strong low-index/ground-state bias in Fig. 2a from T1 leakage, high oracle R at O(10^2) shots does not establish that an unknown molecule can be reconstructed. This is load-bearing for the “practical generative readout” claim.","section":"Section 4.3, Eq. (15)"},{"comment":"Sec. 4.3, Eq. (16): SNR is introduced as a “ground-truth-free proxy,” but it is defined with labeled partitions ⟨N_valid⟩ over the 10 true atoms and ⟨N_empty⟩ over the 182 empties. That measures separability given M*, not recovery without M*. Either redefine SNR from unlabeled histogram statistics only, or drop the “ground-truth-free” language and add an actual unsupervised recovery experiment under the observed hardware bias.","section":"Section 4.3, Eq. (16)"},{"comment":"Sec. 3.1 and App. E: preparation of |ψ_M*⟩ is assumed to exist; the demo uses isometry initialization yielding ~1213-gate 8-qubit circuits, already near coherence-limited depth. The abstract and conclusions frame the scheme as a “practical… representation for molecular geometries on near-term devices” and a building block for QAE/QGAN. Without either a scalable preparation path or a clear scope restriction to “readout of already-prepared sparse support states,” the NISQ practicality claim overreaches. Tighten claims to readout-only, or quantify when preparation remains feasible.","section":"Section 3.1; Appendix E; Abstract/Conclusions"},{"comment":"Sec. 3.1 collision-free grid: the worst-case condition √3 s_voxel < d_min would require s_voxel < 0.63 Å, but experiments use 0.95 Å and rely on molecule-specific orientation/centering. For multi-molecule datasets the paper defers “optimal collision-free grid design.” For the claimed use in generative libraries this is not a minor engineering detail—it determines qubit count and whether the encoding is universal. At minimum, state the qubit/resolution scaling for a fixed universal grid over a standard set (e.g., QM9 subset) or clearly limit claims to single-molecule, orientation-fixed encoding.","section":"Section 3.1; Section 5"}],"minor_comments":[{"comment":"Abstract and Intro compare to O(3^n × 10^{2–3}) full tomography; for fairness also cite classical shadows / Pauli grouping limits when the task is full geometry recovery, as partially done in Sec. 2, so the baseline is not only naive tomography.","section":"Abstract; Section 2"},{"comment":"Fig. 2: mark the 10 true c_a values from Table 1 on the histogram (or an inset) so readers can see whether true atoms sit above the noise floor under hardware bias.","section":"Figure 2"},{"comment":"App. C.4 η̂ estimates (0.19–0.24) are lower than the direct valid-shot fraction ~0.29 in Fig. 3; a one-sentence reconciliation would help.","section":"Appendix C.4"},{"comment":"Typos/notation: “V AEs” spacing in Intro; consistent use of M* vs M^∗; “angstr¨ om” encoding in Sec. 5.","section":"Introduction; Section 5"},{"comment":"Dates say July 30, 2026 / experiments April 2026—fine for arXiv, but confirm consistency before journal submission.","section":"Title page; Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The core coupon-collector readout idea is sound and the Kingston demo is real, but the manuscript sells a generative NISQ building block while only demonstrating oracle recall of a classically prepared support. I would accept a revised version that (i) runs and reports unsupervised support recovery under the measured bias, (ii) corrects the SNR claim, and (iii) scopes preparation honestly. Without (i), the central application claim remains unproven. Fit is reasonable for a quantum-computing / QML venue if claims are tightened; less so if left as a broad “practical molecular design on NISQ” result."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful core is simple and real. They lay a voxelized molecule into computational-basis labels (voxel index × atom type), prepare an equal superposition over the A occupied states, and recover the atoms by sampling. Noise-free that is just the coupon collector: O(A log A) shots. On IBM Kingston they run an 8-qubit ethylamine instance and get mean oracle recall 0.94 at 116 shots and 0.98 at 200, with ~70% wasted shots and clear T1 ground-state bias. Math in 3.2/App. C is standard and correctly applied; the hardware stats (recall curves, wasted-shot breakdown, simulator baseline, Table 1) are cleanly reported. That is a legitimate methods-plus-demo contribution for people building quantum generative models who need a readout-compatible geometry layout.\n\nWhat is actually new is the product encoding plus the support-recovery framing and the Kingston numbers. Basis encoding, voxels, and coupon collector are not new; the package and the experiment are.\n\nSoft spots, in proportion. (1) State preparation is assumed out of scope. The demo uses ~1213-gate isometries that already sit near coherence limits; they say so. End-to-end claims versus tomography are therefore readout-stage only and should stay framed that way. (2) The stress-test point lands: R is an oracle metric (known M*), and their SNR is also computed from labeled valid/empty partitions. Fig. 2a shows strong low-index bias; the true atom indices are scattered. Nothing shows that frequency ranking or thresholding recovers the unknown support rather than decoherence-favored empties. For the generative pipeline they sketch in Sec. 5, that unsupervised step is load-bearing and unproven. (3) Single small molecule, lossy quantization, no rotational invariance, no bonds. Minor relative to the demo’s stated scope. No code shipped.\n\nWho it is for: QML/NISQ people working on molecular generative readout, not chemists needing accurate geometries or anyone expecting a full data-loading solution. The thinking is clear and the literature engagement is honest. I would send it to referees; they should force the oracle-vs-unsupervised distinction and the prep caveat into the abstract and claims. Worth a look if that subfield is on your desk; not a must-read otherwise.","headline":"Solid NISQ encoding-and-demo paper: coupon-collector readout works on hardware for oracle recall, but unsupervised support recovery under real device bias is not shown and state prep is left open.","tokens_in":19034,"tokens_out":593,"would_cite":false,"duration_ms":22475,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Encoding molecules as sparse equal superpositions turns geometry readout into a coupon-collector problem needing only O(A log A) shots.","keywords":["quantum voxel encoding","molecular geometry","NISQ","coupon collector","support recovery","computational-basis sampling","readout efficiency","state tomography alternative"],"falsifier":"Prepare the same 8-qubit ethylamine support state on comparable hardware and check whether mean reconstruction recall stays near 0.94 at ~116 shots and near 0.98 at ~200 shots; failure to reach high recall at those budgets, or circuits so deep that the support is destroyed before sampling, would refute the practical claim.","tokens_in":18879,"feed_emoji":"⚛️","tokens_out":887,"duration_ms":19939,"temperature":0.7,"pith_summary":"Reading a full molecular geometry out of a quantum state usually means expensive tomography that scales exponentially with qubit number. This paper argues that if you discretize space into voxels and put each atom’s position and type into its own computational-basis label, then store the molecule as an equal superposition over those few occupied labels, reconstruction collapses to sampling until every atom appears at least once. In the ideal case that takes only on the order of A log A shots for an A-atom molecule, not exponential tomography. On real IBM hardware the authors reconstruct a 10-atom ethylamine molecule from an 8-qubit circuit with mean recall near 1 using a few hundred shots, even though most shots are noise. The point is a readout-friendly representation for near-term quantum models that already output quantum states, at the cost of spatial quantization to the voxel grid.","feed_headline":"Molecule readout drops to O(A log A) shots via sparse voxels","feed_subtitle":"Equal superpositions over atom-voxel basis states turn geometry recovery into coupon collecting on NISQ hardware","key_machinery":"Sparse quantum voxel encoding: each atom maps to one basis state via a combined voxel-index and atom-type index; the molecule is the equal superposition over those occupied states, so reconstruction is classical coupon collection on measurement outcomes.","core_discovery":"Once a molecule is prepared as an equal superposition over the A computational-basis states that encode its voxelized atom positions and types, complete support recovery by computational-basis sampling requires O(A log A) shots in the noise-free case, and on noisy NISQ hardware a 10-atom molecule can still be recovered at high mean recall with only O(10^2) shots—orders of magnitude fewer measurements than full state tomography.","pith_inferences":["If state-preparation depth cannot be cut well below the demonstrated ~1200-gate regime, the encoding’s readout win may stay confined to proof-of-concept circuits rather than larger generative models.","Adaptive or symmetry-aware grids (hinted in the paper’s outlook) would be the natural next test: same coupon-collector math, less orientation dependence and fewer wasted basis states.","The method is closest in spirit to basis encoding with a deliberately sparse support; neighbouring quantum ML tasks that only need set-like outputs could reuse the same support-recovery pattern."],"forward_implications":["Quantum generative models that emit states in this encoding can hand molecular geometries to classical post-processing with coupon-collector shot budgets instead of tomography.","Measurement cost at readout scales with atom count A, not with 3^n Pauli settings, for fixed voxel resolution.","Spatial precision is capped by voxel size, forcing an explicit qubits-versus-resolution trade-off when targeting sub-ångström work.","Signal-to-noise over empty voxels can serve as a ground-truth-free proxy to decide which basis states are real atoms in a generative setting."],"fun_headline_variants":["Sparse voxels cut molecule readout to O(A log A) shots","Atom-voxel superpositions turn geometry into coupon collecting","10-atom molecule recovered on NISQ with O(100) sparse shots","Support recovery replaces tomography for voxelized molecules","Equal superposition encoding yields readout-efficient geometries"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The paper assumes a usable way already exists to prepare or learn that sparse equal-superposition state; without it the cheap readout does not yield a practical end-to-end method.","fun_headline_variants_meta":{"raw":{"variants":["Sparse voxels cut molecule readout to O(A log A) shots","Atom-voxel superpositions turn geometry into coupon collecting","10-atom molecule recovered on NISQ with O(100) sparse shots","Support recovery replaces tomography for voxelized molecules","Equal superposition encoding yields readout-efficient geometries"]},"model":"grok-4.5","effort":"low","cost_usd":0.003983,"raw_usage":{"total_tokens":1291,"prompt_tokens":832,"num_sources_used":0,"completion_tokens":63,"cost_in_usd_ticks":39828000,"prompt_tokens_details":{"text_tokens":832,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":396,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":832,"tokens_out":63,"duration_ms":7142,"temperature":1.0,"reasoning_tokens":396,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T16:48:36.600681+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Prepare the same 8-qubit ethylamine support state on comparable hardware and check whether mean reconstruction recall stays near 0.94 at ~116 shots and near 0.98 at ~200 shots; failure to reach high recall at those budgets, or circuits so deep that the support is destroyed before sampling, would refute the practical claim.","supporting_citations":[],"review_version":1}