{"id":"67ed546e-4146-4d3f-93f7-f7016cf5e657","arxiv_id":"2504.16350","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A transformer trained on ADAPT-QAOA solutions can generate valid QAOA circuits for unseen MaxCut instances, matching ADAPT-QAOA approximation ratios within about 0.005 while avoiding iterative parameter optimization.","lead":"The paper trains a small GPT model on thousands of ADAPT-QAOA circuits for MaxCut graphs, then uses it to generate near-optimal QAOA circuits for unseen graphs in a single forward pass. It reports approximation ratios close to ADAPT-QAOA on graphs with 10 to 14 nodes, with much faster inference than iterative optimization.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported ARs are normalized by a heuristic MaxCut value rather than exact OPT; for n≤14 exact OPT is cheap, so the paper's central numerical comparison is not anchored to ground truth.","rationale":"I read the paper as a proof-of-concept that a transformer can imitate ADAPT-QAOA's circuit generation. The strongest claim is quantitative: QAOA-GPT achieves mean best AR 0.971–0.972 against ADAPT-QAOA's 0.973–0.974 on unseen MaxCut graphs. The weakest link in that chain is the definition of OPT(G), which is a classical heuristic used both as the training-data filter and as the evaluation denominator. Since exact MaxCut is computationally trivial at n≤14, the authors could have removed this ambiguity at negligible cost; their failure to do so leaves the specific AR values and the reported closeness unverified. The best-of-5 reporting and the omitted training-data cost are real limitations but they do not threaten the existence of a learned graph-to-circuit mapping as directly as a potentially mis-calibrated optimality denominator. The exact-MaxCut check is decisive: it will either confirm the reported AR values or show how much they shift. I agree with the reader's weakest-assumption analysis and the CONDITIONAL verdict, which should remain unchanged pending this verification.","tokens_in":13991,"tokens_out":15112,"duration_ms":151282,"concrete_test":"For each of the 1000 test graphs per problem size (n=10, 12, 14), compute OPT(G) exactly by exhaustive enumeration over all 2^n cuts or by an exact ILP solver. Recompute the mean best AR and mean average AR for QAOA-GPT, ADAPT-QAOA, and the standard QAOA baselines using this exact OPT, and compare with Table I. Also reapply the training-data filter α ≥ 0.97 to the ADAPT-QAOA circuits using exact OPT to see what fraction of the training set actually satisfies the threshold. If the QAOA-GPT-to-ADAPT gap changes by more than 0.003, or if either method's mean AR shifts by more than 0.01, the headline numerical comparison is not robust to the choice of OPT.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III Step 1 defines the approximation ratio as α = ⟨Hc⟩/OPT(G), where OPT(G) is 'the MaxCut value obtained via classical heuristics [25]'. This heuristic value enters the pipeline twice: it filters the training circuits (α ≥ 0.97) and serves as the denominator for every AR reported in Table I and Figure 3. For the tested sizes n=10, 12, and 14, exact MaxCut can be computed by enumerating 2^n cuts in well under a second per graph, yet the authors do not verify against exact values. If the heuristic returns a feasible cut (a lower bound), the reported ARs are inflated and the gap to ADAPT-QAOA is overestimated; if it returns a relaxation upper bound, the reported ARs are depressed. Because both QAOA-GPT and ADAPT-QAOA are divided by the same per-graph heuristic value, the relative ordering is preserved, but the specific headline magnitudes ('0.971 to 0.972', 'within 0.002 to 0.005') are not established. The claim that QAOA-GPT matches ADAPT-QAOA within a few thousandths of the optimum therefore rests on the accuracy of an unverified heuristic.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces QAOA-GPT, a decoder-only transformer trained to autoregressively generate quantum circuits for weighted MaxCut instances. The model conditions on tokenized graph edge lists plus FEATHER graph embeddings and outputs circuit operators and parameters in a single forward pass, bypassing the iterative gradient-based optimization used by ADAPT-QAOA. Training data are generated by ADAPT-QAOA with either the standard QAOA mixer pool or a dual-qubit Pauli pool, and circuits are filtered by an approximation-ratio threshold alpha >= 0.97. Experiments on 1000 unseen Erdős–Rényi graphs per size n=10, 12, 14 report mean best AR of 0.971–0.972, within roughly 0.002–0.005 of ADAPT-QAOA and above standard QAOA by about 0.03. The paper also includes ablations on graph embeddings, training distribution, target AR, operator pool choice, and runtime scalability.","tokens_in":14209,"tokens_out":9121,"duration_ms":86625,"significance":"If the numerical claims are supported, the paper is a useful proof-of-concept that generative transformers can synthesize QAOA-like circuits with quality close to adaptive variational methods while avoiding iterative optimization at inference time. The evaluation against an external approximation-ratio objective, rather than only against training loss, mitigates the most obvious form of supervised-imitation circularity, and the FEATHER-conditioning ablation is informative. The GPU-accelerated data-generation scaling results are also relevant to practical use. The quantitative significance, however, is not yet established: the reported approximation ratios depend on a heuristic MaxCut value that is never checked against exact optima for the small tested sizes, and the best-of-five reporting without per-graph uncertainty makes the headline margins under-specified. No code or data are released, which further limits reproducibility.","major_comments":[{"comment":"The approximation ratio is defined as alpha = <H_c>/OPT(G), where OPT(G) is described as 'the MaxCut value obtained via classical heuristics [25]' (Section III, Step 1). This heuristic value is used twice: it filters the ADAPT-QAOA training circuits at alpha >= 0.97, and it is the denominator of every AR reported in Table I and Figure 3. For the evaluated sizes n=10, 12, and 14, exact MaxCut can be computed by enumerating all 2^n cuts, yet the paper does not verify the heuristic against exact values. If reference [25] returns feasible cuts (a lower bound), all reported ARs are inflated and the gap to ADAPT-QAOA is overstated; if it returns a relaxation upper bound, the ARs are depressed. The headline magnitudes '0.971 to 0.972' and 'within 0.002 to 0.005' are therefore not anchored to ground truth. Please recompute every reported AR using exact OPT for these sizes, and re-examine the training-data filter in light of the exact values.","section":"Section III, Step 1; Table I; Figure 3"},{"comment":"The paper reports QAOA-GPT 'best AR' as the maximum over five generated circuits per graph, but the figures and table present only the mean over graphs with a standard deviation. No confidence intervals on the mean, no distribution over the five generated circuits, and no paired statistical comparison with ADAPT-QAOA are given. Because the claimed margin to ADAPT-QAOA is only 0.002–0.005 in AR, selection noise from best-of-five reporting can affect the stated ranking, and the standard deviation across graphs in Table I does not quantify this. Please report the per-graph best-of-five distribution (mean, worst-of-five, spread across seeds) and a confidence interval or paired test for the ADAPT-QAOA and standard-QAOA comparisons.","section":"Section IV-B; Table I; Figure 3"},{"comment":"The runtime comparison in Figure 7 compares a single forward pass of the trained QAOA-GPT with the entire iterative ADAPT-QAOA optimization, which is appropriate for the paper's speed claim, but the measurement protocol is underspecified: hardware, batch size, number of ADAPT-QAOA layers or optimizer iterations, and any equal-quality criterion are not stated. The text asserts 'exponential scaling' for ADAPT-QAOA, but the figure shows no raw data, error bars, or fitted scaling. Please specify the protocol, report repeated measurements, and clarify what circuit quality is held fixed when comparing inference time with iterative optimization time.","section":"Section V-F; Figure 7"}],"minor_comments":[{"comment":"The sentence 'demonstrate that QAOA-GPT, generates high quality quantum circuits' contains an extra comma after 'QAOA-GPT'.","section":"Abstract"},{"comment":"The opening phrase 'In the this section' has a duplicated article and should read 'In this section'.","section":"Section V"},{"comment":"The table header is ambiguous: the columns labeled 'QAOA-GPT (Best)' and 'QAOA-GPT' under both AR and number of layers do not state which is average and which is best-of-five; Figure 3 distinguishes 'Avg. AR' and 'Best AR', so the table should use matching labels.","section":"Table I"},{"comment":"The text says 'Each data point represents the mean AR over 50 random graphs within the corresponding density bin' and 'A total of 1000 random graphs were evaluated for each problem size.' If densities range from 0.3 to 0.9 in increments of 0.1, seven bins times fifty graphs gives 350 graphs, not 1000; please clarify the binning scheme.","section":"Section IV-B"},{"comment":"There are several typos in this section: 'noisless' should be 'noiseless', 'statevctor' should be 'statevector', and 'scinario' should be 'scenario'.","section":"Section V-G"},{"comment":"The caption 'target = 1 10 4' should be typeset as a mathematical expression, e.g., alpha >= 1 - 10^{-4}, to be readable.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The heuristic-OPT issue is the primary technical blocker and should be the focus of the revision: exact MaxCut values are cheap for n<=14 and can be used to re-anchor all reported ARs. I would also encourage the editors to require release of at least the test graphs, generated circuits, and evaluation scripts, since the paper's claims are empirical and no code or data are currently provided. The paper is better positioned as a proof-of-concept for generative circuit synthesis than as a definitive benchmark; the authors should be asked to strengthen the statistical reporting and the runtime-measurement protocol before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"QAOA-GPT is a serious proof-of-concept: train a decoder-only transformer to synthesize QAOA circuits for MaxCut, conditioned on FEATHER embeddings, and get approximation ratios within a few thousandths of ADAPT-QAOA on small graphs (n=10–14). The combination of graph embeddings, custom tokenization, and GPT training for QAOA circuits is genuinely new relative to GQE and GroverGPT, and the paper does a careful job with ablations—showing that the embeddings matter, that training data diversity helps, and that stricter quality thresholds improve the model. The dual-qubit pool experiment is honest about the model's limits at high precision. Credit where it's due: this is a clean empirical study, not a hype piece.\n\nThe soft spots are real but fixable. The main one, which the stress-test note nails, is that the approximation ratio denominator is OPT(G) from a classical heuristic [25], not exact MaxCut. For n=10, 12, 14, exact values are cheap to compute, so there is no excuse for not anchoring the headline numbers. Because both QAOA-GPT and ADAPT-QAOA are divided by the same per-graph heuristic, the relative comparison probably survives, but the absolute claims like \"0.971 to 0.972\" are not established. That should be fixed with exact OPT before the paper is accepted anywhere.\n\nSecond, the reporting is uneven: QAOA-GPT is evaluated as best-of-5 circuits while the baselines are single-run, and Figure 3 has no error bars. The runtime comparison in Figure 7 is inference-only and omits the cost of generating the training data, which the authors themselves say can take weeks on 500 CPU workers. That's a major amortized cost, and the claim of \"orders-of-magnitude faster circuit generation\" is misleading if it doesn't include the data generation investment. Finally, no code or data are released, which makes it hard to reproduce or build on.\n\nNone of these undermine the central concept, though. The paper is a useful engineering contribution: a demonstration that a GPT can learn to imitate ADAPT-QAOA's output and generalize to new instances. It's not a new quantum algorithm, and it can't beat its teacher, but as a way to avoid per-instance gradient optimization, it's plausible and worth taking seriously.\n\nWho is this for? Researchers working on ML-assisted quantum circuit generation or QAOA variants. It deserves a serious referee—the methodology is mostly sound, the problem is relevant, and the weaknesses are correctable. I'd accept it for peer review but send it back with requests for artifacts and exact-OPT evaluation.","headline":"Solid proof-of-concept for training a GPT to generate QAOA circuits, but the evaluation is not anchored to exact MaxCut and the missing artifacts weaken the reproducibility.","tokens_in":14806,"tokens_out":1629,"would_cite":false,"duration_ms":17670,"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":"A GPT model trained on ADAPT-QAOA circuits can synthesize near-optimal MaxCut circuits in a single forward pass.","keywords":["QAOA","ADAPT-QAOA","generative pretrained transformer","quantum circuit generation","MaxCut","graph embeddings","variational quantum optimization","CUDA-Q"],"falsifier":"Compute exact MaxCut values (by brute force or an exact solver) for the same 1000 test graphs per size used in Table I and recompute the mean best approximation ratios of QAOA-GPT and ADAPT-QAOA; if the gap exceeds the reported 0.002-0.005 or QAOA-GPT's exact AR falls below 0.97, the central claim is refuted.","tokens_in":13751,"feed_emoji":"⚛️","tokens_out":8092,"duration_ms":70151,"temperature":0.7,"pith_summary":"The paper introduces QAOA-GPT, a decoder-only transformer trained to output MaxCut quantum circuits directly from graph inputs. It aims to establish that a single-pass generative model can replace the iterative gradient-based loop of QAOA and ADAPT-QAOA without sacrificing solution quality: on unseen graphs of 10, 12, and 14 nodes, the best sampled circuit achieves mean approximation ratio 0.971-0.972, within 0.002-0.005 of ADAPT-QAOA's 0.973-0.974. Because inference is one forward pass, circuit generation runtime is nearly constant in problem size, in contrast to the steep growth of ADAPT-QAOA. This matters if the bottleneck of variational quantum optimization is the classical outer loop: a learned generator could make circuit synthesis cheap enough for routine use.","feed_headline":"GPT learns to write MaxCut circuits in a single pass","feed_subtitle":"On unseen graphs up to 14 nodes, generated circuits hit 0.971 approximation ratio without gradient optimization.","key_machinery":"The central object is the QAOA-GPT model itself: a decoder-only transformer (a GPT-2-style architecture trained from scratch) that predicts circuit tokens one at a time. Each training instance is a tokenized weighted edge list for a graph followed by tokenized layer blocks of an ADAPT-QAOA circuit, where each block records an operator index and the optimized parameters ($\\gamma_k$, $\\beta_k$). A FEATHER graph embedding (a characteristic-function-based structural descriptor) is broadcast into the input at every token position, letting the model condition circuit generation on global graph structure. The mechanism replaces the variational outer loop: at inference the model samples a full circuit in one forward pass, including both operator identities and parameter values, with no gradient evaluation or classical parameter optimization.","core_discovery":"The paper's discovery, on its own terms, is that a GPT-style transformer trained on ADAPT-QAOA-generated circuits can autoregressively synthesize problem-specific QAOA circuits for new MaxCut instances, jointly producing operator choices and variational parameters. The paper reports that the best of five sampled circuits for each graph tracks ADAPT-QAOA's approximation ratio to within a few thousandths (mean best AR 0.971-0.972 versus 0.973-0.974 across n = 10, 12, 14), while the generated circuits are comparable in layer count. It also reports that graph embeddings are necessary for out-of-distribution generalization, and that raising the target approximation ratio of the training circuits raises the model's output quality. The conclusion the authors draw is that generative inference can stand in for iterative variational optimization when near-optimal solutions are sufficient.","pith_inferences":["The same graph-embedding-plus-edgelist conditioning could extend to other QUBO problems such as graph partitioning, community detection, or coloring, since the cost Hamiltonian enters only through the ADAPT-QAOA training labels.","If inference cost stays nearly constant while ADAPT-QAOA data-generation cost grows, the practical scaling bottleneck shifts to training-data generation; one testable prediction is that model AR should improve predictably with dataset size and diversity, following a scaling law.","Because the model emits both operators and parameters, its outputs could serve as a warm start for ADAPT-QAOA or conventional QAOA, potentially cutting the number of optimizer iterations needed for high precision."],"forward_implications":["For new MaxCut instances, QAOA-GPT generates circuits with no gradient evaluations or classical parameter optimization at inference, so circuit generation time remains nearly constant as problem size grows.","The target approximation ratio used to filter ADAPT-QAOA training circuits directly controls the model's output quality: models trained on circuits with higher target AR produce higher-AR circuits.","Incorporating graph embeddings improves performance on out-of-distribution graph families, with the largest gains on Barabasi-Albert and bipartite graphs.","Training on a mix of graph families (Erdos-Renyi and Barabasi-Albert) improves generalization to Barabasi-Albert graphs without hurting Erdos-Renyi performance.","When the target precision is tightened to $\\alpha \\ge 1 - 10^{-4}$ and a dual-qubit mixer pool is used, QAOA-GPT's approximation ratio plateaus near 0.8, showing that high-precision circuit generation is not yet solved."],"supporting_citations":[{"why":"Defines ADAPT-QAOA, whose gradient-based ansatz construction produces the training circuits and the primary quality baseline.","marker":"[8]"},{"why":"Introduces the QAOA ansatz and alternating cost/mixer structure that QAOA-GPT learns to reproduce.","marker":"[6]"},{"why":"Supplies the classical heuristic used to estimate OPT(G), the denominator in every approximation-ratio label and evaluation.","marker":"[25]"},{"why":"Provides the FEATHER graph embeddings used as structural conditioning in the transformer input.","marker":"[24]"},{"why":"Implements the GPU-accelerated statevector simulation used to generate ADAPT-QAOA training data.","marker":"[13]"},{"why":"Documents the CUDA-Q platform whose GPU simulation speeds up the gradient-evaluation step of dataset generation.","marker":"[14]"},{"why":"Supplies the GPT pretraining objective and decoder-only architecture that QAOA-GPT adapts for circuit sequences.","marker":"[21]"}],"fun_headline_variants":["GPT writes MaxCut circuits without gradient-based search","Autoregressive transformer designs QAOA circuits for unseen graphs","Adaptive QAOA circuits, now generated by GPT in one shot","Generative AI produces compact quantum circuits for MaxCut"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported approximation ratios assume the classical heuristic used for OPT(G) returns the true MaxCut value; if it is inaccurate for the tested graphs, the quality labels on training circuits and the comparison with ADAPT-QAOA are both biased.","fun_headline_variants_meta":{"raw":{"variants":["GPT writes MaxCut circuits without gradient-based search","Autoregressive transformer designs QAOA circuits for unseen graphs","Adaptive QAOA circuits, now generated by GPT in one shot","Generative AI produces compact quantum circuits for MaxCut"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000412,"raw_usage":{"total_tokens":2120,"prompt_tokens":918,"completion_tokens":1202,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":1136}},"tokens_in":534,"tokens_out":1202,"duration_ms":11765,"temperature":1.0,"reasoning_tokens":1136,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:05:54.653680+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute exact MaxCut values (by brute force or an exact solver) for the same 1000 test graphs per size used in Table I and recompute the mean best approximation ratios of QAOA-GPT and ADAPT-QAOA; if the gap exceeds the reported 0.002-0.005 or QAOA-GPT's exact AR falls below 0.97, the central claim is refuted.","supporting_citations":[{"cited_title":"Adaptive quantum approximate optimization algorithm for solving combinato- rial problems on a quantum computer","cited_arxiv_id":null,"evidence_quote":"Defines ADAPT-QAOA, whose gradient-based ansatz construction produces the training circuits and the primary quality baseline."},{"cited_title":"Rank-two relaxation heuristics for max-cut and other binary quadratic programs","cited_arxiv_id":null,"evidence_quote":"Supplies the classical heuristic used to estimate OPT(G), the denominator in every approximation-ratio label and evaluation."},{"cited_title":"Characteristic functions on graphs: Birds of a feather, from statistical descriptors to parametric models","cited_arxiv_id":null,"evidence_quote":"Provides the FEATHER graph embeddings used as structural conditioning in the transformer input."},{"cited_title":"cuquantum sdk: A high-performance library for accelerating quantum science","cited_arxiv_id":null,"evidence_quote":"Implements the GPU-accelerated statevector simulation used to generate ADAPT-QAOA training data."},{"cited_title":"NVIDIA CUDA-Q framework","cited_arxiv_id":null,"evidence_quote":"Documents the CUDA-Q platform whose GPU simulation speeds up the gradient-evaluation step of dataset generation."}],"review_version":1}