{"id":"29467fc2-2262-47ff-8b2d-a1853d617d71","arxiv_id":"2504.14459","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A machine learning method learns a pure quantum state's classical description using only SWAP-test fidelity as feedback, demonstrated in simulation and on single-qubit IBM hardware, but it is framed misleadingly as a non-destructive snapshot.","lead":"This paper trains a machine learning model to reconstruct a quantum state from a similarity score (SWAP test fidelity), storing the state's classical description for later reload. The authors claim this is a non-destructive quantum snapshot, but the method requires repeated re-preparation of multiple copies, so it does not actually observe a single state without destroying it.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The protocol's own copy requirement contradicts the advertised non-destructive single-copy snapshot claim; the method is a copy-consuming, tomography-like estimator, not a state-preserving memory.","rationale":"The reader's weakest_assumption correctly identifies the repeated-copy requirement as the load-bearing flaw. My independent reading confirms it: the paper's Introduction explicitly distinguishes itself from QST by claiming no multiple copies are needed, but Section III-A and Algorithm 2 require repeated reinitialization. A SWAP test between a candidate and an unknown state does not leave the unknown copy untouched; the controlled-swap entangles it with the ancilla, and the protocol resets all qubits after each fidelity estimate. Consequently, every fidelity number reported — in simulation, noisy simulation, and hardware — is a statistical estimate obtained from many copies of a state whose preparation is known to the experimenter. This makes the abstract's claim of observing states 'without direct measurement, preserving them for reuse' false as stated. The method's reframed value, if any, is as a machine-learning approach to pure-state tomography using SWAP-test fidelity as a loss signal; that is a legitimate but much more modest claim, and the paper does not clearly restrict itself to it. I agree with the reader's REJECT verdict: the central claim is internally inconsistent with the protocol, and no experiment or code is provided that would demonstrate the advertised single-copy, non-destructive memory operation. The proposed copy-count audit would settle the issue cleanly and could be reported in a revised manuscript.","tokens_in":15702,"tokens_out":5191,"duration_ms":54772,"concrete_test":"Instrument Algorithms 1 and 2 with a copy counter that increments every time the unknown state is freshly reinitialized before a fidelity evaluation, as required by Section III-A. Run the same 1-qubit hardware experiment and a 6-qubit noiseless simulation, and report total copies of the unknown state consumed per reconstructed state. If the counter is greater than 1 — for QESwap it will be at least 50 times the number of iterations — then the claimed 'preserving the state for reuse' and 'single-copy, non-destructive' properties are not realized by the submitted protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central advertised capability — observing a quantum state without destroying it, preserving it for later reuse, and doing so at multiple points within a running circuit — is explicitly contradicted by the paper's own protocol. Section I states that QST 'cannot serve as a mechanism for capturing quantum snapshots, where the goal is to capture the state non-destructively and without requiring multiple copies.' Yet Section III-A states: 'We repeatedly reinitialize the unknown quantum state, thereby requiring multiple copies, an assumption aligned with the principles of quantum state tomography.' Section III-D adds: 'States are re-prepared in each iteration.' Meanwhile Section III-C describes QESwap as 'designed for single-copy, non-destructive quantum learning,' but Algorithm 2 evaluates a population of N=50 candidates per iteration, and each fidelity evaluation requires a fresh instance of the unknown state because the SWAP test disturbs or consumes the copy. For up to 100 iterations, one reconstruction consumes at least thousands of copies, not one. This is not a minor caveat: it invalidates the 'non-destructive snapshot' and 'classical quantum memory' framing. What is actually demonstrated is a classical optimization loop that learns a pure-state description from many re-prepared copies, i.e., a variant of quantum state tomography. The no-cloning theorem makes the advertised single-copy preservation impossible in general; the paper's own text admits the workaround. The mixed-state limitation, acknowledged in Section V-B, further narrows the method to pure states. Because the headline claim and the protocol's copy budget are mutually inconsistent, the central result as stated is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a machine-learning-driven method, named QESwap, for reconstructing unknown pure quantum states using only SWAP-test-based fidelity feedback. The authors combine a deep-neural-network generator and an evolutionary-strategy optimizer with iterative SWAP tests, and report near-unit fidelities in noiseless and noisy simulations for up to six qubits, as well as single-qubit demonstrations on IBM hardware. The central claim is that this enables non-destructive 'quantum snapshots' of states at arbitrary points in a circuit, with classical storage and later reuse, thereby functioning as a form of quantum memory. The paper also provides comparative benchmarks, an entanglement-entropy analysis of reconstructed states, and a detailed gate-level noise model.","tokens_in":15930,"tokens_out":2548,"duration_ms":26620,"significance":"If the central claim were correct, the ability to observe a quantum state non-destructively, store it classically, and reuse it would be a major advance with implications for quantum debugging, QRAM, and modular circuit design. The paper has some genuine strengths: it openly discusses the mixed-state limitation of the SWAP test, provides a realistic noise model, and reports reproducible-looking simulation benchmarks over many random states. However, the advertised capability is not supported by the protocol as written. The method repeatedly re-prepares the unknown state and consumes copies at every fidelity evaluation, so it is a copy-consuming, tomography-like estimation scheme rather than a single-copy non-destructive snapshot. The no-cloning theorem makes the advertised general capability impossible, and this is not a peripheral caveat but a direct contradiction of the paper's main contribution. Consequently, the significance of the reported numerical results is confined to a classical optimization loop over many identical copies, which is a much weaker and less novel claim.","major_comments":[{"comment":"The protocol contradicts the paper's central claim of non-destructive observation. Section III-A states that the authors 'repeatedly reinitialize the unknown quantum state, thereby requiring multiple copies, an assumption aligned with the principles of quantum state tomography,' and Section III-B2(c) says each fidelity evaluation uses 'a re-prepared instance of the unknown target state.' The SWAP test itself disturbs or consumes the copy used for the overlap estimate. Therefore the method does not observe a state without destroying it; it estimates a state from many fresh copies, exactly like quantum state tomography. This invalidates the headline claims of 'non-destructive snapshot' and 'single-copy, non-destructive quantum learning' made in Sections I and III-C.","section":"Section III-A and III-B2(c)"},{"comment":"The copy consumption is quantitatively severe. Algorithm 2 evaluates a population of N=50 candidates per iteration, and each evaluation requires a separate freshly prepared instance of the unknown state. With up to 100 iterations, a single reconstruction consumes thousands of copies of the target state. The reader should not mistake this for a resource-light observation: this is a classical optimization loop over multiple copies, and the claimed equivalence to a non-destructive 'classical quantum memory' is not established by the reported experiments.","section":"Algorithm 2 and Section III-C"},{"comment":"The claim of observing states 'at different points within a single circuit, one at a time' is not supported. Section III-D acknowledges that 'states are re-prepared in each iteration.' A snapshot at a circuit point would require the state at that point to be preserved or at least accessible without re-preparation; instead, the method produces a statistical estimate from many independent runs. This does not enable mid-circuit introspection of a single evolving computation, and the QRAM, debugging, and modular-program applications in Section V-A therefore rest on an unsupported capability.","section":"Section III-D"},{"comment":"The hardware validation is limited to three known single-qubit states (|0>, |1>, and (|0>+|1>)/sqrt(2)) on ibm_sherbrooke. No unknown or hidden states are used, no multi-qubit states are tested, and the fidelity on hardware is not compared against any baseline. This is a useful sanity check for the simulation pipeline, but it does not support the paper's broad claims about capturing arbitrary quantum states or about scalability to the 6-qubit systems reported in simulation.","section":"Section IV, Real Hardware Results"}],"minor_comments":[{"comment":"The notation is inconsistent: Section III-B1 defines the unknown target as |ψ> while Algorithm 1 and the surrounding text call it |ϕ_u>. Please pick one symbol and use it consistently.","section":"Section III-B1 and Algorithm 1"},{"comment":"The text says the bit-flip probability of 2.003e-04 represents a '0.1% chance of bit flips,' but 2.003e-04 is 0.02%. Please correct the percentage or the value.","section":"Appendix A"},{"comment":"The bound P_success ≤ 3/4 for SWAP-test-based discrimination of two non-orthogonal mixed states is stated without a proof or reference. Please provide a citation or a derivation.","section":"Section V-B"},{"comment":"Several table entries are marked 'NA' for noisy simulations (e.g., gradient-based unitary matrix, 2 and 3 qubits). The figure caption and text interpret these as cases that failed to converge, but the table itself does not say so explicitly. Please add a footnote defining 'NA' consistently.","section":"Table III"},{"comment":"The final fidelity for all simulated experiments is computed classically via the inner product with the known target state. This is fine as a validation step, but the text should be clearer that this evaluation uses the classically known target vector, not the SWAP-test estimate, so the reported fidelity is not the noisy hardware fidelity in general.","section":"Section IV, Benchmarking"}],"recommendation":"reject","confidential_remarks":"The manuscript's central claim is contradicted by its own protocol: the method consumes multiple copies and is essentially a tomography-like estimator, not a non-destructive snapshot. This is a load-bearing conceptual error that cannot be fixed by local revisions within the current scope. The numerical results may be of interest as a benchmark for classical optimization over SWAP-test fidelities, but the paper as written would need a fundamentally reframed contribution and likely new experiments to support any non-destructive claim. I would not encourage resubmission in its present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this one. The headline claim—non-destructive observation of quantum states and classical storage without consuming copies—is contradicted by the paper's own protocol. And once you strip that claim away, there's a modest, plausibly useful tomography variant underneath.\n\nWhat's actually new: using SWAP-test fidelity as the only learning signal, with both a neural generator and an evolutionary strategy (QESwap), and no pre-collected training dataset. That combination appears to be absent from the cited ML-QST works, which use projective measurement datasets. The authors also report a small hardware demo on IBM, which is more than many such papers do.\n\nThe soft spot is load-bearing. Section III-A explicitly says the unknown state is repeatedly reinitialized, requiring multiple copies. Section III-D says states are re-prepared in each iteration. So the 'single-copy, non-destructive snapshot' language in the abstract and introduction is not supported. This matters because the entire motivation—observing states at arbitrary circuit points without destruction—depends on it. SWAP tests also disturb the states they compare, so a single copy can't be reused indefinitely anyway. The honest description is: quantum state tomography with a fidelity oracle instead of projective measurements. That is a legitimate thing to study, but not the thing the paper claims.\n\nThe mixed-state limitation is acknowledged and actually discussed well—they notice the SWAP test estimates Hilbert-Schmidt inner product, not true fidelity, for mixed states. That's a point in their favor.\n\nWhere it's thin: hardware validation is three known single-qubit states, no error bars, no code or data released. The final evaluation computes fidelity classically via inner product, so the 'fidelity-only' claim applies to training, not verification. The population size and noise scale are manually set; no sensitivity analysis. Minor point: the comparison table in the intro is a bit self-serving, but not factually wrong.\n\nSo the paper is not nonsense. It's an overclaimed incremental contribution. A reader interested in learning-to-tomograph from similarity oracles might find the ES approach worth a look, but I wouldn't cite the 'quantum memory' framing.\n\nVerdict: deserves a serious referee, but with a clear instruction to reframe. If judged on the advertised capability, it's a reject. If judged on the underlying protocol, it's a weak accept after major revision. The authors need to either drop the single-copy language or show a real single-copy protocol, which no-cloning probably forbids.","headline":"A plausibly useful tomography variant buried under a non-destructive-snapshot claim that the protocol itself contradicts.","tokens_in":16525,"tokens_out":2605,"would_cite":false,"duration_ms":24749,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","81P45"],"pacs":["03.67.-a"],"model":"deepseek-v4-flash","headline":"The paper argues that an unknown pure quantum state can be reconstructed, stored classically, and recreated on demand through a guess-and-check loop whose only feedback is a SWAP-test fidelity number, with no direct measurement of the…","keywords":["quantum snapshot","SWAP test","fidelity estimation","evolutionary strategy","quantum state reconstruction","quantum memory","QRAM","NISQ hardware"],"falsifier":"Prepare an uncharacterized state that has no preparation circuit, run the loop once, and ask whether any stored vector reproduces it; it cannot, because the SWAP test needs freshly re-prepared copies every round, so no fidelity signal ever exists. A quantitative version: reconstruct a two-qubit Bell state on real superconducting hardware using fidelity-only feedback and compare the result with the claimed 0.99 threshold, since the paper's hardware evidence covers only single-qubit states and its own noisy simulations show fidelity slipping as the qubit count rises.","tokens_in":1861,"feed_emoji":"⚛","tokens_out":2310,"duration_ms":96518,"temperature":0.7,"pith_summary":"This paper claims that an unknown pure quantum state can be learned, and later recreated, without ever measuring the state directly. The method runs a guess-and-check loop: a classical generator proposes a candidate state, a SWAP test compares the candidate against a freshly re-prepared copy of the unknown state and returns a single fidelity number, and that number is the only feedback the generator receives as it refines its guess. Once the fidelity approaches 1, the winning candidate is stored as a classical vector and can be reloaded into any compatible circuit, which the authors present as a working form of classical quantum memory useful for debugging, circuit introspection, and QRAM-style reuse. In noiseless simulation the loop reaches average fidelity 0.999 over 100 random states, and on real superconducting qubit hardware it reconstructs single-qubit states at near-unit fidelity. The authors themselves delimit the claim: the method requires many re-preparations of the unknown state and does not extend to mixed states, because the SWAP test returns the Hilbert-Schmidt inner product rather than true fidelity.","feed_headline":"SWAP tests rebuild quantum states at 0.999 fidelity","feed_subtitle":"A guess-and-check loop that never measures states directly stores them as classical vectors.","key_machinery":"The load-bearing mechanism is the SWAP test used as a non-destructive fidelity oracle. A candidate state and a re-prepared copy of the unknown state are loaded side by side with an ancilla qubit, and the expected value $\\langle Z\\rangle$ on the ancilla equals the squared overlap $|\\langle\\psi|\\varphi\\rangle|^2$, so a single measurement statistic feeds the entire learning loop without ever measuring either state directly. The second mechanism is a classical generator whose output is a complex vector, normalized to unit norm and converted into a quantum circuit through a uniformly-controlled-rotation state-preparation routine, so that each classical guess can be physically instantiated and compared. The third is the evolutionary update rule of QESwap, $w \\leftarrow w + \\frac{\\alpha}{N\\sigma}\\sum_i A_i z_i$, in which Gaussian noise vectors $z_i$ generate a population of candidates, the SWAP test scores each one, and standardized advantages $A_i$ reweight the update; this derivative-free loop is what keeps working under noise and on hardware.","core_discovery":"On its own terms, the paper's claim is that non-destructive observation of a quantum state reduces to maximizing a black-box fidelity function. The SWAP test provides the oracle: for two states on $n$ qubits, an ancilla-based comparison yields $\\langle Z\\rangle = |\\langle\\psi|\\varphi\\rangle|^2$, so the expected Pauli-$Z$ value on the ancilla is exactly the overlap between the candidate and the unknown state. Feeding that single number back into a generator turns reconstruction into an optimization problem in which neither state is ever measured directly. The paper demonstrates two optimizers: a deep neural network trained with finite-difference gradients through the non-differentiable SWAP test, and a gradient-free evolutionary strategy (QESwap) that perturbs a population of candidate state vectors and updates them with standardized fidelity rewards. Both reach fidelity above 0.99 within roughly 5 to 17 epochs in noiseless simulation; under noise the evolutionary strategy is the more stable of the two, and it is the variant demonstrated on real hardware, converging on single-qubit known states to near-unit fidelity within three epochs. The paper also draws its own boundary: density-matrix reconstruction of mixed states plateaus near 0.8 fidelity because the SWAP test's Hilbert-Schmidt inner product overestimates mixed-state similarity, so pure states are the intended scope.","pith_inferences":["The snapshot is a batch property, not a single-instance capture: the loop re-prepares the unknown state many times, so a state that appears only once in a circuit cannot be snapshotted, which quietly limits the debugging story to states that can be regenerated from a known preparation procedure.","What is stored is ultimately a preparation recipe rather than a quantum memory in the information-theoretic sense; the classical vector is only as good as the ability to re-prepare the state it describes, an ability the method assumes but does not certify.","A natural testable extension is to replace the SWAP-test oracle with an overlap estimator that is well-defined for mixed states, such as randomized-measurement or shadow-based fidelity; the paper's own analysis pinpoints the Hilbert-Schmidt inner product as the blocker, so the fix is localized.","If the same fidelity-only loop scales to multi-qubit entangled states on real hardware, the claimed memory architecture becomes substantially more credible, but the paper's hardware evidence covers only three single-qubit states, so scaling is the open risk."],"forward_implications":["Intermediate quantum states in a circuit can be captured one at a time, stored as classical vectors, and re-prepared later, giving near-term devices a form of non-volatile memory bank without persistent physical qubits.","Because the stored state is a plain classical vector, it can be retrieved by standard QRAM access schemes and fed into downstream circuits, enabling modular program design where whole subcircuits need not be re-executed.","The fidelity-only feedback signal means no training dataset and no access to the unknown density matrix are required, so the loop can run directly against noisy hardware; the paper reports near-unit hardware fidelity for single-qubit states.","Mixed states are outside the method's scope: the SWAP-test signal saturates at a success probability of 3/4 for distinguishing non-orthogonal mixed states, so the framework is limited to pure-state snapshots.","The gradient-free evolutionary strategy is the recommended configuration, reaching fidelity above 0.99 in noisy simulation where the gradient-based network stalls or fails."],"supporting_citations":[{"why":"States the exponential (3^n measurement settings) cost of standard quantum state tomography that the snapshot method aims to bypass.","marker":"[5]"},{"why":"Defines QRAM, the retrieval setting into which the classically stored reconstructed states are meant to plug.","marker":"[13]"},{"why":"Supplies the uniformly-controlled-rotation state-preparation method used to turn each guessed classical vector into a quantum circuit for comparison.","marker":"[21]"},{"why":"Provides the state-preparation primitive and simulation platform the framework's implementation is built on.","marker":"[22]"},{"why":"Gives the evolutionary-strategy update rule that QESwap adapts to SWAP-test fidelity feedback.","marker":"[23]"}],"fun_headline_variants":["SWAP test snapshots: quantum memory without direct measurement","Guess, SWAP, repeat: capture quantum states at 0.999 fidelity","Machine learning stores quantum states in classical memory via SWAP","Non-destructive quantum observation: SWAP test enables state reuse","Quantum introspection at 0.999 fidelity with SWAP test"],"cache_read_input_tokens":18560,"weakest_assumption_plain":"The loop assumes the unknown state can be prepared over and over again from a known procedure; if only one copy exists, there is nothing to SWAP-test against, and no snapshot is possible.","fun_headline_variants_meta":{"raw":{"variants":["SWAP test snapshots: quantum memory without direct measurement","Guess, SWAP, repeat: capture quantum states at 0.999 fidelity","Machine learning stores quantum states in classical memory via SWAP","Non-destructive quantum observation: SWAP test enables state reuse","Quantum introspection at 0.999 fidelity with SWAP test"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000402,"raw_usage":{"total_tokens":2161,"prompt_tokens":1077,"completion_tokens":1084,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":693,"completion_tokens_details":{"reasoning_tokens":996}},"tokens_in":693,"tokens_out":1084,"duration_ms":9417,"temperature":1.0,"reasoning_tokens":996,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:48:05.975259+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Prepare an uncharacterized state that has no preparation circuit, run the loop once, and ask whether any stored vector reproduces it; it cannot, because the SWAP test needs freshly re-prepared copies every round, so no fidelity signal ever exists. A quantitative version: reconstruct a two-qubit Bell state on real superconducting hardware using fidelity-only feedback and compare the result with the claimed 0.99 threshold, since the paper's hardware evidence covers only single-qubit states and its own noisy simulations show fidelity slipping as the qubit count rises.","supporting_citations":[{"cited_title":"Pennylane: Automatic differentiation of hybrid quantum-classical computations,","cited_arxiv_id":null,"evidence_quote":"Provides the state-preparation primitive and simulation platform the framework's implementation is built on."}],"review_version":1}