{"id":"44c6ceb6-cf8f-47c8-8524-576d90660914","arxiv_id":"2501.03518","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A quantum annealing solver for combinatorial optimization can be trained on a classical simulator and then transferred to real quantum hardware, improving convergence and speed.","lead":"This paper shows that a trainable combinatorial optimization solver, whose internal step sizes are normally learned on a classical computer, can then run on a quantum annealer and still solve problems quickly. The idea is to train the solver's parameters with a classical simulation of quantum annealing, then transfer those parameters to real quantum hardware.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The transfer premise is asserted but never calibrated: D-Wave samples are not compared with Q(x;v), so the SQA-trained step sizes could be brittle under hardware noise and effective-temperature mismatch.","rationale":"The reader's weakest assumption is exactly the one I find most load-bearing: SQA-trained step sizes must remain effective when the sampler is swapped for D-Wave. I agree with that assessment. The transfer premise is not tested by any calibration of D-Wave samples against Q(x;v); Figs. 4 and 5 only show that SQA is a better classical surrogate than MH and that larger Trotter numbers help. The failure mode is concrete: the gradient in Eq. (8) and the stability factor in Eq. (7) depend on the quantitative value of <f_k>; a hardware bias in these expectations changes the effective update and can in principle make the learned eta_t unstable. I also note a secondary issue: Eq. (6) uses a diagonal Jacobian in the chain rule, whereas covariances between different f_k create off-diagonal terms; this concerns the training rule rather than the transfer claim and does not change the conditional verdict. The timing comparison in Table I is internally consistent and is not the main weak point. The missing distributional calibration and the single-image scope both warrant a conditional accept; the proposed test, comparing empirical constraint moments from D-Wave and SQA at fixed v and then checking whether matching the target distribution changes the result, would settle whether the transfer premise lands. Since the concern is a missing calibration rather than a known contradiction, I would keep the reader's CONDITIONAL verdict rather than moving to rejection.","tokens_in":10084,"tokens_out":8872,"duration_ms":88483,"concrete_test":"Calibrate the D-Wave sampler against the theoretical Q(x;v(t)): fix the trained SQA-QA trajectory, and at 5-10 values of t keep v(t) fixed, generate 10,000 samples each from SQA and D-Wave for the QUBO representing Q(x;v(t)), and compare the empirical means <f_k(x)> and covariances with very-long-run SQA (e.g., Trotter number 8) estimates. Report the maximum normalized bias and total-variation distance. Then re-run the SQA-QA MSE and timing experiments after tuning D-Wave anneal time and/or post-processing so that the measured <f_k> match Q(x;v) to within SQA sampling error. If the MSE-zero-by-30 result persists, the transfer premise is robust; if it changes materially, the claimed benefit is an artifact of an uncalibrated hardware distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism is the transfer premise stated in Sec. IV: step sizes {eta_t} trained with a classical sampler are used with a quantum annealer 'if the difference is sufficiently small.' The paper's only support is indirect: SQA-QA beats MH-QA in Fig. 4, and larger Trotter numbers improve Fig. 5. There is no direct measurement of how close D-Wave's output distribution is to Q(x;v(t)) = exp(-beta f0(x) + beta sum_k v_k f_k(x))/Z, the distribution whose expectations drive update (4) and whose variance enters the training gradient (8). D-Wave samples from an annealed transverse-field Ising system with an unknown effective temperature, noise, and anneal-time-dependent nonequilibrium effects; these are not identical to SQA's path-integral target. If <f_k>_QA is biased relative to <f_k>_Q, Eq. (4) tracks the wrong constraint expectations, and the stability factor 1 - eta_u ∂<f_k>/∂v_k in Eq. (7) can turn negative, causing divergence. The headline 'MSE reaches zero within 30 iterations' is an average over 50 instances of one fixed 15x15 image, and the paper reports no QA sampling parameters (anneal time, reads, embedding, post-processing). Thus the observed success may reflect QA's near-ground-state sampling on this small sparse problem rather than a validated transfer of learned parameters.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes classical-quantum transfer learning for the deep-unfolded Ohzeki method (DUOM). Step sizes of the Ohzeki gradient-descent updates are trained entirely with a classical sampler -- either Metropolis-Hastings (MH) or simulated quantum annealing (SQA) -- and the trained step sizes are then transferred to DUOM executed on a D-Wave quantum annealer. The method is tested on a 15x15 binary image reconstruction problem with N=225 linear measurements and M=135 constraints. Numerical results show that SQA-trained DUOM executed on QA (SQA-QA) reaches zero MSE within 30 iterations, outperforms MH-trained transfer and a fixed-step-size grid-search baseline, and has lower total execution time than SQA-SQA. The paper also studies the effect of the SQA Trotter number and concludes that SQA is a more suitable classical surrogate than MH.","tokens_in":10410,"tokens_out":4210,"duration_ms":42061,"significance":"If the transfer premise is valid, the proposal is practically valuable: it removes quantum computation from the training loop while retaining quantum hardware at inference, and it extends deep unfolding to quantum samplers. The paper is commendably concrete: it reports error bars, compares multiple samplers, includes a grid-search baseline, and evaluates execution time, not just iterations. The SQA-vs-MH comparison is a sensible way to probe surrogate quality, and the claim that SQA-trained parameters transfer better than MH-trained parameters is an informative empirical result. However, the central validation rests on a single fixed ground-truth image and on an uncalibrated assumption that D-Wave sampling statistics are close enough to the SQA target distribution; both need to be addressed before the broader conclusions can be accepted.","major_comments":[{"comment":"The transfer premise is asserted but never directly calibrated. Section IV states that parameters trained with a classical sampler can be used with a quantum annealer if the difference is sufficiently small, but the manuscript provides no measurement of the difference between D-Wave samples and the theoretical distribution Q(x; v(t)) used in the update rule (4) and the stability factor (7). If <f_k>_QA is biased relative to <f_k>_Q, the step sizes trained under SQA may be suboptimal or cause divergence. The paper should either compare empirical constraint expectations <f_k> obtained from D-Wave and SQA at the trained v(t) trajectory, or report a sensitivity analysis over D-Wave sampling parameters (anneal time, number of reads, chain strength, postprocessing), none of which are currently reported.","section":"Section IV and Section V.B"},{"comment":"All experiments use one fixed ground-truth image x* (Fig. 3), with the dataset consisting of random matrices A and corresponding observations y = A x*. Thus the headline 'MSE reaches zero within 30 iterations' is an average over 50 instances of a single image. This limits the support for the general claim that the learned step sizes accelerate convergence and for the conclusion's statement of 'practical viability'. The authors should test on multiple independent ground-truth images or otherwise justify that the learned step sizes are not image-specific.","section":"Section V.A and Figures 4-6"},{"comment":"The Trotter-number experiment is confounded by the simultaneous scaling of the inverse temperature: the figure reports results for (beta, tau) = (1,1), (4,4), and (8,8). Because beta itself controls the sharpness of Q(x;v), the improved performance with larger tau cannot be attributed to a better approximation of QA dynamics; it may simply reflect a lower-temperature target distribution. The authors should fix beta (or vary it systematically) while changing the Trotter number to isolate the effect of the SQA approximation.","section":"Section V.B.2 and Figure 5"},{"comment":"The execution-time comparison is not fully reproducible as reported. The manuscript does not state the number of D-Wave API calls (presumably one per iteration, so 30 calls for SQA-QA), the per-call anneal time, the number of reads, or the embedding parameters that produce the reported 181 ms sampling time. The 12.2 s API column dominates the SQA-QA total, so the '190 times faster' statement applies only when API latency is ignored, while the practical speed-up is 2.9x. The authors should report the API call structure and hardware settings so that the wall-clock comparison can be reproduced and interpreted fairly.","section":"Section V.B.3 and Table I"}],"minor_comments":[{"comment":"There is a duplicated phrase 'without without QA during the training process'; please revise.","section":"Section IV"},{"comment":"The text contains several typos ('data-riven', 'e fficient', 'Ohezeki') that should be corrected in a final pass.","section":"Section III.B"},{"comment":"The loss function is denoted L_lambda in Eq. (2) and the text, but Eq. (6) uses L without a subscript; please make the notation consistent.","section":"Equation (6) and (7)"},{"comment":"The sentence 'SQA-QA achieved the highest performance' should be qualified, since Figure 6 shows that SQA-SQA converges faster than SQA-QA; the claim is true only among the QA-executed methods compared in Figure 4.","section":"Section V.B.1"},{"comment":"No code or data availability statement is included. Given that the paper is empirical and relies on specific random instances and hardware settings, providing code and the exact instance generation parameters would substantially improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the uncalibrated transfer premise does land: the paper's central mechanism is only supported indirectly. The Trotter-number confound and the single fixed image are additional load-bearing gaps, but they are all addressable with additional experiments or clarified scope. I would not reject; the core idea is plausible and the experimental methodology is mostly sound. I would condition acceptance on the calibration and confounding issues being resolved, and on the authors being more careful about the generality of their claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Phil,\n\nTwo things to know. First, the idea is genuinely practical: train the step sizes of the deep-unfolded Ohzeki method (DUOM) on a classical sampler, then swap in a quantum annealer at execution time, avoiding the high cost of QA during training. Second, the headline result holds for what they tested: SQA-trained step sizes transfer to the D-Wave machine and reach zero MSE within 30 iterations on a 15x15 image reconstruction problem, and the execution time is about 190x faster than SQA on a CPU when API latency is excluded.\n\nWhat's new: this is the first application of classical-quantum transfer learning to the Ohzeki-style solver. The empirical comparison between SQA and MH as training samplers is useful—SQA transfers better, which is expected but nice to see quantified. The Trotter-number dependence is a sensible sanity check, and the execution-time table is honestly presented, with and without API overhead.\n\nThe soft spots are real but not fatal. The empirical scope is narrow: one problem, one fixed ground-truth image, 50 random measurement matrices. No code or data, and no D-Wave sampling parameters (anneal time, reads, embedding) are reported, so reproducibility is limited. The transfer premise—that learned parameters survive swapping SQA for QA—is asserted rather than directly calibrated. They never compare D-Wave output to the theoretical Q(x;v), so we don't know how aligned the samplers actually are. The stress-test concern is legitimate: the success on this small, sparse problem may partly reflect QA's near-ground-state sampling, and the transferred step sizes could be brittle on harder instances. That said, the empirical result is still evidence the approach can work; it just doesn't tell us how far it generalizes.\n\nThe conclusion overreaches a bit with \"practical viability\"—I'd call it a promising proof-of-concept. But the central claim holds for what they actually tested. This is a reasonable extension of the authors' prior DUOM work, and it deserves a serious referee. I'd send it to review, but ask for more experimental detail and at least one more problem class before acceptance.","headline":"Classical-quantum transfer learning for the deep-unfolded Ohzeki method is a practical idea that works on the single benchmark they tried, but the evidence base is too narrow to support the conclusion's stronger claims.","tokens_in":10925,"tokens_out":2907,"would_cite":false,"duration_ms":27875,"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":"SQA-trained step sizes transferred to a quantum annealer reach zero error within 30 iterations and cut total solving time from 34.9 s to 12.2 s.","keywords":["quantum annealing","deep unfolding","Ohzeki method","classical-quantum transfer learning","combinatorial optimization","simulated quantum annealing","image reconstruction","step-size learning"],"falsifier":"Compare the empirical distribution of samples from the quantum annealer with SQA samples at the same $Q(x;v)$; if the difference in the estimated constraint expectations $\\langle f_k\\rangle$ is comparable to or larger than the learned step size $\\eta_t$ times the gradient scale, the transferred parameters can be suboptimal or diverge. A direct version is to run SQA-QA on instances with $M/N > 0.633$ and check whether MSE still reaches zero within 30 iterations; failure would show that the transfer result depends on the easy regime.","tokens_in":9908,"feed_emoji":"⚛️","tokens_out":9314,"duration_ms":80811,"temperature":0.7,"pith_summary":"The paper tries to establish that a trainable quantum-annealing-based combinatorial optimization solver can be trained entirely on a classical computer and still deliver a speed advantage on real quantum hardware. It proposes classical-quantum transfer learning for the deep-unfolded Ohzeki method (DUOM), in which step sizes learned with a classical sampler are transferred to a quantum annealer. On a binary image reconstruction task, the SQA-trained version executed on a quantum annealer reached zero reconstruction error within 30 iterations and took 12.2 s total, versus 34.9 s for the fully classical SQA-trained DUOM. If correct, this removes the main obstacle to using quantum annealers inside trainable solvers: the prohibitive cost of training on quantum hardware.","feed_headline":"Trained quantum annealer solver cuts runtime from 34.9 s to 12.2 s","feed_subtitle":"Step sizes learned on a classical simulator transfer to real annealing hardware, reaching zero error within 30 iterations.","key_machinery":"The load-bearing object is the deep-unfolded Ohzeki method (DUOM), a trainable version of the Ohzeki method in which the Hubbard-Stratonovich transformation converts a constrained QUBO into repeated sampling from $Q(x;v) \\propto \\exp(-\\beta f_0(x) + \\beta \\sum_k v_k f_k(x))$ followed by gradient updates $v_k \\leftarrow v_k + \\eta_t(C_k - \\langle f_k\\rangle)$. Deep unfolding makes the step sizes $\\eta_t$ learnable parameters, and the variance identity $\\partial\\langle f_k\\rangle/\\partial v_k = \\beta(\\langle f_k^2\\rangle - \\langle f_k\\rangle^2)$ supplies gradients without differentiating through the sampler. Classical-quantum transfer learning is the scheme of training these $\\eta_t$ with a classical sampler (MH or SQA) and then executing the same unfolded iterations with a quantum annealer; the Trotter number in SQA controls how closely the training sampler approximates QA.","core_discovery":"The central claim is that step sizes trained under simulated quantum annealing (SQA) are good enough, after transfer, for DUOM executed on a quantum annealer to solve constrained binary quadratic problems faster and more accurately than the untrained Ohzeki method with a grid-searched step size. In the experiments SQA-QA achieved the best mean squared error (MSE), reaching zero within 30 iterations, while the untrained solver did not reach zero within 50; SQA-QA also beat MH-QA, indicating that the classical sampler used for training matters. In wall-clock terms, the transferred solver took 12.2 s per instance including API latency, while the classically executed SQA-SQA took 34.9 s, even though SQA-SQA converged in fewer iterations. This is the paper's evidence that classical-quantum transfer learning can make a trainable quantum COP solver practical.","pith_inferences":["The reported speed advantage depends on the quantum annealer's sampling rate and API latency; on a different device or with local access, the wall-clock comparison could shift, and the paper's 12.2 s figure is dominated by API response time.","The transfer's validity rests on an untested distributional match between SQA and the hardware annealer; a systematic comparison of hardware samples with $Q(x;v)$ could select the best classical surrogate and Trotter number instead of assuming SQA is close.","The experiments use one fixed $15\\times 15$ image with $M/N = 0.6$ below the $0.633$ threshold; testing near or above that threshold, or on inequality-constrained knapsack instances, would show whether the zero-MSE result is a property of the method or of the easy regime.","The same recipe—train an unfolded iterative algorithm on a classical surrogate, then deploy it on quantum hardware—should apply to other trainable quantum-classical solvers beyond DUOM, provided the surrogate's output distribution is close enough to the hardware's."],"forward_implications":["SQA-trained step sizes transferred to a quantum annealer reach zero MSE within 30 iterations on the tested image reconstruction instances, while the untrained Ohzeki method with constant step size does not reach zero within 50.","Larger Trotter numbers in SQA improve transfer: $\\tau=4$ and $\\tau=8$ succeed where $\\tau=1$ (equivalent to MH at $\\beta=1$) fails.","Executing DUOM on a quantum annealer reduces total wall-clock time to 12.2 s per instance versus 34.9 s for the classically executed SQA-SQA, even though SQA-SQA converges in fewer iterations.","Because DUOM accepts any sampler, the training phase can be kept classical, avoiding the roughly one million sampling calls per training run that quantum execution would require."],"supporting_citations":[{"why":"Supplies the Ohzeki method, the sampling-based constrained COP solver whose step sizes DUOM learns.","marker":"[10]"},{"why":"Introduces DUOM and the variance-based gradient estimation that makes step-size training possible.","marker":"[13]"},{"why":"Provides the Hubbard-Stratonovich transformation that linearizes the penalty terms and defines the sampled distribution $Q(x;v)$.","marker":"[11, 12]"},{"why":"Establishes SQA as the classical emulator of quantum annealing used for training and transfer.","marker":"[25, 26]"},{"why":"Provides the optimizer used to update the learned step sizes during training.","marker":"[31]"},{"why":"Supplies the simulated quantum annealing sampler implementation used in the training experiments.","marker":"[39]"},{"why":"Sets the statistical-mechanical threshold used to place the test problem in a hard underdetermined regime.","marker":"[37]"},{"why":"Identifies the quantum annealer hardware used in the execution phase of SQA-QA.","marker":"[5]"}],"fun_headline_variants":["Transfer learning speeds up quantum annealer solver 3x","Classical training accelerates quantum combinatorial solver","Hybrid transfer learning cuts quantum solver runtime","Trained step sizes make quantum annealer converge in 30 steps","Quantum solver with transfer learning reaches zero error faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The step sizes trained under SQA remain effective when the sampler is swapped for a real quantum annealer, which requires the hardware annealer's output distribution to be close enough to SQA's at the distributions $Q(x;v)$ encountered during execution.","fun_headline_variants_meta":{"raw":{"variants":["Transfer learning speeds up quantum annealer solver 3x","Classical training accelerates quantum combinatorial solver","Hybrid transfer learning cuts quantum solver runtime","Trained step sizes make quantum annealer converge in 30 steps","Quantum solver with transfer learning reaches zero error faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1387,"prompt_tokens":888,"completion_tokens":499,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":425}},"tokens_in":504,"tokens_out":499,"duration_ms":5200,"temperature":1.0,"reasoning_tokens":425,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:52:20.187038+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the empirical distribution of samples from the quantum annealer with SQA samples at the same $Q(x;v)$; if the difference in the estimated constraint expectations $\\langle f_k\\rangle$ is comparable to or larger than the learned step size $\\eta_t$ times the gradient scale, the transferred parameters can be suboptimal or diverge. A direct version is to run SQA-QA on instances with $M/N > 0.633$ and check whether MSE still reaches zero within 30 iterations; failure would show that the transfer result depends on the easy regime.","supporting_citations":[{"cited_title":"Choi, Quant","cited_arxiv_id":null,"evidence_quote":"Supplies the Ohzeki method, the sampling-based constrained COP solver whose step sizes DUOM learns."},{"cited_title":"Ohzeki, Sci","cited_arxiv_id":null,"evidence_quote":"Introduces DUOM and the variance-based gradient estimation that makes step-size training possible."},{"cited_title":"Monga, Y","cited_arxiv_id":null,"evidence_quote":"Provides the optimizer used to update the learned step sizes during training."},{"cited_title":"Evenbly and G","cited_arxiv_id":null,"evidence_quote":"Supplies the simulated quantum annealing sampler implementation used in the training experiments."},{"cited_title":"Wierichs, J","cited_arxiv_id":null,"evidence_quote":"Sets the statistical-mechanical threshold used to place the test problem in a hard underdetermined regime."}],"review_version":1}