{"id":"cd4a2f38-1c58-4410-862e-b25a56f691ba","arxiv_id":"2501.16776","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper introduces the Heat Exchange (HE) ansatz, a variational circuit built from XX+YY interactions with bath qubits, and reports improved MaxCut approximations and sub-1% Heisenberg ground-state errors in small numerical tests.","lead":"Researchers propose a new building block for variational quantum circuits, inspired by algorithmic cooling, that links each problem qubit to an auxiliary qubit through an excitation-swap interaction. They report better MaxCut solutions than QAOA on small random graphs and under 1% error on a six-spin Heisenberg chain with an impurity, but the evidence is limited.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The MaxCut superiority claim rests on an ansatz that reduces to a classical product-state search; without a classical baseline the claim is unsubstantiated.","rationale":"I read the paper as claiming an incremental ansatz design rather than a fault-tolerant quantum advantage, and the dVQE/Heisenberg portion is separate and may be reproducible from the text. The reader's conditional verdict already flags missing expressibility analysis and missing classical baselines. My stress-test sharpens that concern: for the MaxCut application, the Fig. 2(a) ansatz is provably product-state over the problem qubits and classically simulable, so the central 'superiority' claim needs a classical comparison to have force. I would not reject the paper outright: the claim could be repaired by providing a classical baseline and by narrowing the wording to 'competitive on small instances'. The internal inconsistencies noted by the reader, such as the swap-angle and critical-field issues, remain but do not by themselves falsify the dVQE result. The verdict therefore stays CONDITIONAL, equivalent to UNCHANGED relative to the reader's assessment.","tokens_in":12544,"tokens_out":6717,"duration_ms":69654,"concrete_test":"Re-run the MaxCut experiments of Fig. 4 on the identical random weighted complete graphs with a classical baseline: optimize the cut objective directly over bitstrings using simulated annealing or multi-start greedy local search, using the same number of objective evaluations as the HE ansatz. If the classical baseline matches or beats the reported HE approximation ratio, the HE superiority claim is not supported. A complementary check is to classically optimize the product-state expected cut over the theta_i; if its optimum equals the best deterministic cut for every positive-weight instance, the ansatz family adds no expressiveness beyond classical mean field.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing issue is in Section 2.3 and Fig. 2(a). Each problem qubit Qp interacts only with its private bath qubit Qb through one XX+YY gate, and there are no gates between problem qubits. Starting from a tensor product |0...0>_p |1...1>_b, the circuit is a tensor product of independent two-qubit unitaries, so after tracing out the bath qubits the reduced state on the problem register is a product state. For the diagonal MaxCut cost of Eq. (6), expectation values factor as <Zi Zj> = <Zi><Zj>, so optimizing the angles theta_i is equivalent to a classical product/probability search over bitstrings. The paper does not compare against any classical bitstring optimizer or mean-field baseline on the same instances and budget. Thus the claimed 'superior approximation ratios' over QAOA and HEA in Fig. 4 could be a property of many evaluations of a simple product ansatz on n=5 and n=10, not a property of the HE ansatz as a quantum state family. This is not an external-consensus objection; it follows directly from the circuit structure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a 'Heat Exchange' (HE) variational ansatz built from XX+YY interactions between problem qubits and ancilla 'bath' qubits initialized in |1>, without resets. It applies the ansatz in two settings: (i) weighted MAXCUT on complete graphs with n=5 and n=10, comparing against QAOA and a hardware-efficient ansatz in ideal simulation and on ibm_strasbourg; and (ii) a dissipative VQE variant for the 1D XXX Heisenberg chain with an impurity, comparing ground-state energies against an effective-Hamiltonian reference. The paper claims superior approximation ratios for MAXCUT and sub-1% energy errors for the Heisenberg chain, including an 'edge effect' simulation.","tokens_in":12754,"tokens_out":11854,"duration_ms":106675,"significance":"If the MAXCUT claim were supported, a single low-depth XX+YY block would outperform standard variational circuits on small weighted complete graphs. However, the circuit in Fig. 2(a) prepares a product state on the problem register, so the MAXCUT experiment is equivalent to a classical product-distribution search; the claimed superiority is therefore not a result about a quantum state family and needs a classical baseline before it can be interpreted. The Heisenberg dVQE application is potentially more interesting, with a concrete physical target and a falsifiable edge-effect prediction, but as presented it lacks the specification and statistical detail needed to verify the sub-1% claim. The paper is clearly written in its circuit definitions and uses standard tools (Qiskit, TREX readout mitigation, Gaussian-process plus ImFil optimization), but no code or data are released and the displayed curves carry no error bars.","major_comments":[{"comment":"In the MaxCut ansatz of Fig. 2(a), each problem qubit Qp interacts only with its own bath qubit Qb through one XX+YY gate and there are no entangling gates between problem qubits. Since the initial state is a tensor product and every gate is a tensor product of two-qubit unitaries, the reduced state on the problem register is a product state for all parameters. For the diagonal Hamiltonian of Eq. (6), <Z_i Z_j> = <Z_i><Z_j>, so optimizing the HE angles is exactly equivalent to a classical product-distribution search over bitstrings. The 'superior approximation ratios' reported in Fig. 4 are therefore not evidence about a quantum state family; they can be reproduced by any classical mean-field bitstring optimizer. The manuscript must include a classical baseline (e.g., local search, simulated annealing, or direct product-distribution optimization) on the same instances with the same evaluation budget, and the MaxCut claims must be reframed accordingly.","section":"Section 2.3, Fig. 2(a), Eq. (6)"},{"comment":"The Heisenberg analysis contains an internal inconsistency in a load-bearing quantity. The text states that the critical field is h_c = 2J and sets J = 1, which gives h_c = 2, but later reads 'the magnitude of the magnetic field exceeds h_c = 4' and uses h = 4 as the separating field in Fig. 6. The phase assignments and the 'edge effect' interpretation depend on the correct threshold, so this must be corrected and the energies at h = 3 and h = 4 re-examined. In addition, the circuit in Fig. 5 shows only one HE block on qubit 0, while results are reported for impurities at distances d = 1, 2 and for impurity states both |0> and |1>; the manuscript does not state how the impurity is moved or how a |1> impurity is encoded with a cooling block that drives population toward |0>. Without these details the edge-effect simulation is not reproducible.","section":"Section 4.2, Eq. (8), Fig. 6"},{"comment":"The experimental comparison is not apples-to-apples. The real-device curve labeled CHEreal in Fig. 4 is compared against ideal noiseless simulation curves for HEA and QAOA, so the statement that it 'still outperformed ideal simulation result' is biased by the absence of noise in the baselines. Moreover, no error bars, number of independent runs, or significance tests are reported for any of the displayed approximation-ratio or probability curves; with only one displayed instance per size, the claimed ordering of algorithms may not be stable. The manuscript should report statistics over many random graphs and, for the hardware point, compare with the same circuits executed under the same noise conditions or a calibrated noise model.","section":"Section 4.1, Fig. 4"},{"comment":"The role of the HE cooling block in the Heisenberg result is not isolated. The comparison in Fig. 6 is only between the HE-dVQE output and an exact reference energy; there is no ablation against a standard VQE using the same RealAmplitude U block without the HE block, nor against the original dVQE of Ref. [YNMF20]. Since a 6-qubit Heisenberg chain is small enough that the U block alone may already reach sub-1% error, the reader cannot conclude that the cooling mechanism contributes to the reported accuracy. An ablation is necessary to support the causal claim in the abstract and Section 5.","section":"Section 4.2, Fig. 2(b), Fig. 6"}],"minor_comments":[{"comment":"There are several typos and grammatical errors: 'ansatzs' and 'anstaz' should be 'ansatze', and 'fluctutation', 'predefiend', and 'conneced' should be corrected.","section":"Throughout"},{"comment":"The claim that uniform weights in [0,1] 'encompass the Sherrington-Kirkpatrick model' is inaccurate, since SK uses zero-mean, unit-variance couplings; either state the mean-field shift explicitly or remove the claim.","section":"Section 4.1"},{"comment":"The optimization budget is described only as 'maximum iteration number will be a control parameter'; define the budget variable used on the x-axis of Fig. 4 and state the number of independent restarts per curve.","section":"Section 3"},{"comment":"The caption 'typical histogram of circuit count of QAOA' is unclear; specify what is being histogrammed (measurement outcomes, cut values, or circuit counts) and for which parameter values.","section":"Fig. 3(b)"},{"comment":"The phrase 'quantum integer problems' appears to be a typo and should be clarified, for example as 'Ising-type problems'.","section":"Section 2.3"},{"comment":"The bath is said to be 'initialized as the ground state' in Section 2.2, but Fig. 2(a) prepares Qb in |1>; if |1> is the excited state, the cooling direction and sign convention should be stated explicitly, since this affects the interpretation of the HE block.","section":"Section 2.2 and Fig. 2(a)"},{"comment":"Eq. (9) defines tau^k_j as 'Pauli k operators for the i-th bath'; the index should be j, and the raising and lowering operators sigma^± should be defined before use.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The MaxCut section is the main weakness; the circuit is classically simulable, so the abstract overclaims. I would not oppose a resubmission if the authors add classical baselines, statistical rigor, and a corrected, fully specified Heisenberg study. The paper might be better suited to a specialized quantum-optimization or quantum-simulation venue after these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThe paper is a mixed bag. The genuinely new piece is the HE ansatz for dVQE: using an XX+YY interaction to a bath qubit as a coherent cooling block, without resets, to emulate dissipation in a variational circuit for a Heisenberg chain with an impurity. On that task the reported sub-1% errors are plausible and the edge-effect simulation is a nice piece of empirical work, though I would want error bars and more than one instance before trusting the numbers.\n\nThe MaxCut half is the soft spot, and it is load-bearing. As the stress-test note says, the circuit in Fig. 2(a) is a tensor product of independent two-qubit XX+YY gates between each problem qubit and a dedicated bath qubit. With no gates between problem qubits, tracing out the baths leaves a product state. For the diagonal MaxCut cost, correlators factor, so the whole optimization is a classical product-state (mean-field) search. The paper compares this against QAOA and a hardware-efficient ansatz on n=5 and n=10, but never against a classical bitstring search or even a simple mean-field optimizer. So the claimed superior approximation ratios are likely an artifact of comparing a cheap classical search to restricted quantum circuits, not a property of the HE ansatz as a quantum state family. This needs a classical baseline and a fairer budget before any quantum claim can be made.\n\nThere are also smaller internal issues: the swap angle at π/4 is not a complete swap (that happens at π/2), the critical-field description of the Heisenberg chain is sloppy, and Eq. (6) has a sign convention that conflicts with the later approximation-ratio definitions. The real-device MaxCut result is compared against ideal simulations of the competitors, which is not an even comparison.\n\nNone of this looks like fraud; it is missing rigor and overstatement. The dVQE idea is worth taking seriously, but the paper as written overclaims on MaxCut. A revision that adds a classical baseline, error bars, a fair real-hardware comparison, and corrects the internal errors would be worth reading.\n\nMy recommendation: send to peer review. The dVQE part deserves referee time, and the MaxCut flaw is substantive but fixable; a good referee would catch it quickly.","headline":"A clever dVQE cooling block overshadowed by an unsubstantiated MaxCut claim that reduces to classical mean-field search.","tokens_in":13305,"tokens_out":5452,"would_cite":false,"duration_ms":46350,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.Ac"],"model":"deepseek-v4-flash","headline":"The paper claims that a reset-free XX+YY cooling block, applied one-to-one between problem and bath qubits, outperforms QAOA and hardware-efficient ansätze on complete-graph MaxCut, and that the same block inside a dissipative VQE reaches…","keywords":["algorithmic cooling","variational quantum eigensolver","heat-exchange ansatz","MaxCut","Heisenberg chain with impurity","dissipative VQE","NISQ devices","XX+YY interaction"],"falsifier":"The most direct test is to take a random weighted complete graph on 10 nodes, solve MaxCut exactly by brute force, and run the Fig. 2(a) HE ansatz, QAOA at $p=2$, and a classical single-bit-flip local search under identical budgets. Because the HE circuit becomes a product distribution over the problem qubits once the bath qubits are traced out, any instance where the best HE cut is strictly below the QAOA or classical cut would refute the paper's MaxCut superiority claim.","tokens_in":12362,"feed_emoji":"🧊","tokens_out":8965,"duration_ms":75880,"temperature":0.7,"pith_summary":"Most variational ansätze either rely on deep circuits or on resets to a thermal bath to remove entropy. This paper proposes a minimalistic alternative: a Heat-Exchange (HE) block, a single parameterized $XX+YY$ rotation between each problem qubit and an auxiliary qubit prepared in $|1\\rangle$, which coherently transfers population without any bath reset. The authors claim this block alone outperforms QAOA and hardware-efficient ansätze on weighted complete-graph MaxCut in both simulation and noisy-device experiments, and that when the block is used as the dissipative layer of a dissipative VQE it estimates the ground-state energy of a 1D Heisenberg chain with a frozen impurity to below 1% error, reproducing the known edge effect. The point of the claim is practical: if true, a fixed low-depth unitary, implementable with native two-qubit gates, is enough to handle problems that usually force deep problem-specific circuits. That would lower the hardware cost of variational algorithms for optimization and for disordered quantum systems.","feed_headline":"One XX+YY block tops QAOA on MaxCut and spin chains","feed_subtitle":"A reset-free cooling-inspired ansatz reaches sub-1% Heisenberg ground-state errors and higher MaxCut ratios.","key_machinery":"The load-bearing object is the heat-exchange interaction $H^{(j)}_{\\rm int}=J(\\sigma^x_j\\tau^x_j+\\sigma^y_j\\tau^y_j)$ between problem qubit $j$ and a bath qubit, which at $Jt=\\pi/4$ acts as a partial iSWAP swapping populations of $|01\\rangle$ and $|10\\rangle$. The HE ansatz places one such gate between each problem qubit and a bath qubit initialized in $|1\\rangle$, so each gate pushes the problem qubit toward $|0\\rangle$; in the dVQE variant, the same block is inserted at the impurity site as a parameterized cooling layer. The mechanism does the work of population redistribution and decoherence emulation that other approaches achieve with mid-circuit resets or a physical thermal bath. This single two-qubit gate is what carries the entire argument: no reset, no external bath, only a coherent exchange between two qubits in the same circuit.","core_discovery":"On the paper's own terms, the central discovery is that the unitary evolution of heat-bath algorithmic cooling—an $XX+YY$ exchange between a target qubit and a bath qubit that starts in $|1\\rangle$—can be promoted from a cooling primitive into a variational ansatz. The resulting HE block shifts population toward the ground state of each problem qubit through a partial-swap rotation, and it does so without resetting the bath or coupling to an external environment. The paper reports that, used directly as the whole ansatz for complete-graph weighted MaxCut, this block reaches higher approximation ratios and higher probabilities of finding the optimal cut than hardware-efficient and QAOA circuits, including on a real device after readout-error mitigation. Used as the dissipative component of dVQE, the same block yields ground-state energies for a six-site 1D XXX Heisenberg chain with one impurity at distance $d=0,1,2$ from the edge that agree with exact reference values to within 1%, and it reproduces the edge effect in which an edge impurity effectively shortens the chain.","pith_inferences":["Beyond the paper: after tracing the bath qubits, the Fig. 2(a) MaxCut circuit generates only product states, so its MaxCut search is effectively a classical mean-field randomized cut; a comparison against a semidefinite-programming rounding bound would show how much of the reported advantage is due to the circuit rather than to the optimization of independent qubit biases.","Beyond the paper: the same cooling block should transfer to other diagonal-cost problems, such as weighted Max-2-SAT or biased spin-glass instances, where single-qubit partition biases are known to help; a low-cost test is to run the block on 20-node instances with skewed weights.","Beyond the paper: the 1% Heisenberg claim is demonstrated only for six sites; scaling the cooling block to 10–20 sites and checking the error beyond exact diagonalization would show whether the edge-effect physics survives in a regime where classical references are harder to obtain."],"forward_implications":["The HE block gives a fixed, shallow, reset-free replacement for entropy-removal operations in variational circuits, lowering the hardware overhead of VQE on near-term devices.","On weighted complete graphs, including Sherrington–Kirkpatrick-type instances, the one-to-one cooling circuit reaches higher approximation ratios than QAOA at comparable or greater depth.","Inserted as the dissipation block of dVQE, the same block computes impurity-site ground states of 1D Heisenberg chains to within 1% and reproduces the boundary edge effect.","Since the required gate is a partial iSWAP, the ansatz maps directly onto the native two-qubit interactions of superconducting processors.","The recipe generalizes: any algorithmic-cooling operation can be converted into a variational layer when the target problem benefits from controllable population bias."],"supporting_citations":[{"why":"Supplies the original variational-quantum-eigensolver framework into which the HE ansatz is inserted.","marker":"[PMS+14]"},{"why":"Provides the heat-bath algorithmic cooling background that the HE mechanism is drawn from.","marker":"[PRBF+16]"},{"why":"Supports the claim that cooling can be optimized through thermalization strategies, the family the HE block approximates without resets.","marker":"[ALP19]"},{"why":"Defines dissipative VQE, the algorithm whose dissipation block the HE block replaces in the many-body application.","marker":"[YNMF20]"},{"why":"Introduces QAOA, the main baseline the HE ansatz is compared against on MaxCut.","marker":"[FGG14]"},{"why":"Introduces the hardware-efficient ansatz used as the second baseline and as the U block in the dVQE circuit.","marker":"[KMT+17]"},{"why":"Supplies the Gaussian-process surrogate plus ImFil optimizer used in all variational optimizations.","marker":"[MLIdJ22]"},{"why":"Provides the readout-error mitigation technique applied to the real-device MaxCut results.","marker":"[vdBMT22]"},{"why":"Context for deterministic non-unitary dissipative operations inside VQE, which the HE cooling block is designed to simplify.","marker":"[CLB+24]"}],"fun_headline_variants":["Cooling-inspired block beats QAOA on MaxCut and Heisenberg","Reset-free ansatz from cooling tops QAOA on MaxCut","HE ansatz: one block for better MaxCut and spin ground states","No bath reset needed: cooling ansatz wins on MaxCut and chains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The MaxCut comparison assumes that the family of circuits in Fig. 2(a), where every problem qubit interacts only with its own always-|1\\rangle bath qubit through one XX+YY rotation, is expressive enough to reach near-optimal cuts on every random weighted complete graph; after tracing the baths this family reduces to independent single-qubit distributions, so the question is whether that classical mean-field search is actually that good.","fun_headline_variants_meta":{"raw":{"variants":["Cooling-inspired block beats QAOA on MaxCut and Heisenberg","Reset-free ansatz from cooling tops QAOA on MaxCut","HE ansatz: one block for better MaxCut and spin ground states","No bath reset needed: cooling ansatz wins on MaxCut and chains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1558,"prompt_tokens":939,"completion_tokens":619,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":543}},"tokens_in":555,"tokens_out":619,"duration_ms":5992,"temperature":1.0,"reasoning_tokens":543,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T10:48:55.557148+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The most direct test is to take a random weighted complete graph on 10 nodes, solve MaxCut exactly by brute force, and run the Fig. 2(a) HE ansatz, QAOA at $p=2$, and a classical single-bit-flip local search under identical budgets. Because the HE circuit becomes a product distribution over the problem qubits once the bath qubits are traced out, any instance where the best HE cut is strictly below the QAOA or classical cut would refute the paper's MaxCut superiority claim.","supporting_citations":[],"review_version":1}