{"id":"c549f212-0a8e-4f5c-b8f9-5c9d3f1229ed","arxiv_id":"2504.17561","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Evolutionary algorithms with periodic angle optimization reduced the depth of random 4- and 6-qubit circuits by up to roughly 90% while keeping moderate to high fidelity.","lead":"This paper tests evolutionary algorithms that search for short quantum circuits while keeping the output state close to a target. It reports depth reductions of up to about 90% on random 4- and 6-qubit circuits, but with only four random seeds and no comparison to standard compilers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Four-seed averaging without per-circuit variance leaves the headline 80% depth reduction unsubstantiated; reported 6-qubit fidelities (0.71–0.77) also strain the 'high fidelity' qualifier.","rationale":"Good-faith reading: the paper is a clear, honest empirical study. It compares a hybrid EA, a plain EA, a no-EA baseline, and a random baseline, and it reports in the discussion that the hybrid is not always better and that tradeoffs exist. That transparency is a genuine strength. The central weakness is not internal inconsistency but insufficient evidence for the strength of the abstract's claim. The four-seed protocol is underpowered, and the absence of per-seed data means the reader cannot check whether the 80% figure is a stable property of the algorithm or an artifact of a few favorable random circuits. The paper also makes no claim to optimality, so the depth reductions should be read as 'the algorithm found shorter circuits,' not 'these are minimal.' The 6-qubit fidelities of 0.71–0.77 further weaken the 'high fidelity' phrase; the abstract should either restrict the claim to the 4-qubit regime or set an explicit fidelity threshold. These are addressable empirical gaps rather than fatal flaws, so the reader's CONDITIONAL verdict is appropriate and I would not change it. The concrete checks above would either confirm the headline or force a more modest, better-scoped claim.","tokens_in":12290,"tokens_out":7401,"duration_ms":74104,"concrete_test":"Request the authors' code and data (or reimplement the Section 4 algorithm) and rerun the four 4-qubit configurations with at least 30 independently sampled random circuits per configuration, reporting per-circuit depth reductions and 95% confidence intervals. Separately, for the 6-qubit depth-33 configuration, report the maximum fidelity achievable at high (≥80%) depth reduction. If the lower confidence bound for mean depth reduction at depth 20 drops below about 50%, or if no 6-qubit circuit reaches F ≥ 0.9 at ≥80% reduction, the abstract's 'significant' and 'high fidelity' wording should be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, as stated in the abstract, is that the proposed methods 'significantly reduce the depth of circuits while still retaining a high fidelity to the target state.' The numerical anchor is the Fig. 5a result: mean depth reduction 80%, mean fidelity 0.98354. The load-bearing assumption is that this average is representative of behavior on random circuits. Section 5 reports only that 'Experiments were run for 4 different seeds and the plots show the mean of the aggregated results.' It does not state whether the four seeds generate four different random circuits, four runs on one circuit, or a mix, and no per-seed values, error bars, or significance tests accompany the depth/fidelity bars. The same figures show high variability across configurations: the hybrid's depth reduction ranges from 26.6% (4 qubits, depth 47) to 88.64% (6 qubits, depth 23). The 6-qubit depth-33 experiment achieves 88.64%/90.91% depth reduction but with fidelities of only 0.77/0.71, which are hard to characterize as 'high fidelity.' Without a stated sampling protocol or variance information, the headline number is a point estimate of unknown stability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a hybrid evolutionary algorithm for quantum circuit construction and optimization, targeting two objectives: minimize circuit depth and maximize fidelity to a target state. Two EA variants are compared: a hybrid variant that periodically calls a COBYLA parameter-optimization subroutine, and a regular EA without it. Each variant is tested in two modes: initializing the population randomly (construction from scratch) and initializing with the target circuit (existing-circuit optimization). Experiments are run on random 4- and 6-qubit circuits of various depths, with four seeds. The reported results show large depth reductions in some settings (e.g., 80% at fidelity 0.98 for a 4-qubit depth-20 circuit) but also fidelities around 0.71–0.77 for the 6-qubit depth-33 experiments. The paper concludes that the proposed methods significantly reduce depth while retaining high fidelity.","tokens_in":12558,"tokens_out":6063,"duration_ms":53510,"significance":"The paper's main value is its detailed and clearly described evolutionary framework and the systematic ablation of the parameter-optimization subroutine. If the reported effects are stable, the work provides a useful baseline and modular building blocks for quantum architecture search. However, the strength of the central claim is currently limited by the small number of seeds, the absence of variance reporting for depth and fidelity, the lack of standard compiler or synthesis baselines, and the fact that the regular EA often outperforms the hybrid in depth reduction. The manuscript also does not supply code or data, so the point estimates cannot be independently verified.","major_comments":[{"comment":"The central empirical claim is supported only by means over four seeds, with no error bars, per-seed values, or significance tests for depth or fidelity. The text in Section 5 says 'Experiments were run for 4 different seeds' but does not say whether each seed draws a new random circuit or only a new initialization on one circuit. Given the large spread across configurations (from 26.6% to 90.91% depth reduction), the headline 80% figure in Fig. 5a is a point estimate of unknown stability. Please report the sampling protocol and add variance information or statistical comparisons.","section":"§5, Figs. 5–7"},{"comment":"The abstract's 'high fidelity' qualifier is not supported by the 6-qubit depth-33 from-scratch results: the hybrid and regular EAs reach fidelities of only 0.77329 and 0.71449 while reducing depth by 88.64% and 90.91%, respectively. These values are below what is usually meant by high fidelity and are also much lower than the 0.94–0.98 fidelities reported for most other configurations. The claims should be restricted to settings where high fidelity is actually achieved, or the fidelity–depth trade-off at 6 qubits should be discussed explicitly.","section":"§6.1, Fig. 7b"},{"comment":"In most from-scratch experiments, the regular EA without parameter optimization achieves a larger depth reduction than the hybrid EA (e.g., Fig. 5b: 75.0% vs 53.12%; Fig. 5c: 82.43% vs 72.3%; Fig. 5d: 80.32% vs 26.6%; Fig. 7b: 90.91% vs 88.64%). Thus the hybrid variant's advantage is limited to fidelity, and even that does not hold in Fig. 6 (depth-47 without solution optimization, where the 'No EA Op.' baseline reaches 0.98316 fidelity). The paper should characterize the conditions under which the hybrid variant is actually preferable.","section":"§6.1, Figs. 5b–5d and 7b"},{"comment":"The experiments compare the proposed algorithms only against the internal 'No EA Op.' baseline and, in Fig. 8b, a random baseline. There is no comparison with standard quantum circuit optimization or synthesis tools, such as Qiskit's transpiler or established compilers. Consequently, 'significant' depth reduction is demonstrated only with respect to the original randomly generated circuit, not with respect to the methods a practitioner would otherwise use; the practical relevance claim in the introduction and conclusion is therefore under-supported.","section":"§5, §6.1"}],"minor_comments":[{"comment":"The optimizer is attributed to 'scikit-kit learn [22]'; COBYLA is provided by SciPy's optimize module, not scikit-learn. The name and the reference should be corrected.","section":"§4.5"},{"comment":"The sentence 'where d is the depth of the target circuit and 1 the minimum depth' should read 'and 1 is the minimum depth'; also, δ should be defined explicitly before Eq. (5).","section":"§4.7, Eq. (6)"},{"comment":"The random baseline is mentioned in the setup but does not appear in the depth/fidelity figures; it is only visible for fitness in Fig. 8b. Either include it in the main comparisons or explain its absence.","section":"§5, Figs. 5–7"},{"comment":"The depth bars have no indication of variance; Fig. 8 reports standard deviation only for fitness, not for depth or fidelity. Please add error bars or per-seed scatter to the main metric figures.","section":"§6.1, Figs. 5–7"},{"comment":"The manuscript does not provide a code or data availability statement, which would be helpful for reproducibility given the small sample sizes and the absence of error bars.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely acceptable for a venue like GECCO after revision if the authors add variance statistics, clarify the circuit sampling protocol, and temper the 'high fidelity' claim to match the 6-qubit results. If the editorial standard requires stronger empirical validation, the missing baselines and statistical tests could be decisive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper is a clean, incremental contribution to evolutionary quantum circuit optimization. The new piece is the hybrid loop: every 25 generations, COBYLA tunes rotation parameters on a random 10% of the population, while the EA keeps mutating and crossing over circuit structure. That periodic coupling is a reasonable extension of the authors' earlier GA4QCO and related hybrid feature-map work, and the paper gives it a fair first test. The writing is clear and the ablation design is sensible: they compare hybrid vs plain EA vs a no-EA baseline, and they test both construction from scratch and optimization of a given circuit.\n\nWhat's weak is the evidence under the claims. The whole experimental basis is four seeds per configuration, with no error bars, no significance tests, and no per-seed spread. Figure 5 shows wild variation across depths: the hybrid's depth reduction swings from 26.6% to 80%, and at 6 qubits the 'high fidelity' claim sits on fidelities of 0.71–0.77, which are not high by any normal standard. The abstract says 'significantly reduce the depth while still retaining a high fidelity' – that's too strong for the 6-qubit results. Also, there's no comparison to an off-the-shelf compiler (e.g., Qiskit transpile with optimization level 3) or to any established synthesis baseline, so we don't know whether the method beats what a practitioner would already use. They don't release code or data, which makes the four-seed averages even harder to assess.\n\nThe stress-test note about the missing sampling protocol is right. I also think the periodic COBYLA subroutine is doing less than the plain EA in terms of depth reduction in several configurations; the paper's own discussion admits this, but the abstract doesn't. That said, the central qualitative result – that EAs can find shorter circuits with decent fidelity on these small random instances – is plausible, and the paper does not hide its limitations in the discussion.\n\nWho's this for? People working on evolutionary quantum architecture search or circuit synthesis. It's a conference-level incremental result. It deserves peer review rather than desk rejection, but a serious referee should require more seeds, error bars, compiler baselines, and a toned-down abstract before acceptance. If the authors can't provide those, it's not ready for a main track.","headline":"A clean, incremental EA-for-circuit-depth paper whose four-seed experiments don't support the abstract's 'high fidelity' claim, but the hybrid idea and ablations are worth a referee's time.","tokens_in":13067,"tokens_out":2424,"would_cite":false,"duration_ms":23138,"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 hybrid evolutionary algorithm constructs and compresses quantum circuits, cutting depth by 80–90% on random 4- and 6-qubit circuits while keeping fidelity near the target state.","keywords":["Evolutionary Algorithm","Quantum Circuit Construction","Quantum Architecture Search","Quantum Circuit Optimization","Multi-Objective Optimization","Hybrid Evolutionary Algorithm","Circuit Depth","Fidelity"],"falsifier":"Run the same experiments on a larger set of seeds, say 50 or 100, and on structured benchmark circuits such as QFT, GHZ preparation, or QAOA ansätze; if the mean depth reduction on 4-qubit depth-20 circuits falls well below 80% at comparable fidelity, or if some benchmark circuits show no depth reduction at all, the central claim would be weakened.","tokens_in":12116,"feed_emoji":"🧬","tokens_out":5087,"duration_ms":45693,"temperature":0.7,"pith_summary":"This paper argues that a multi-objective evolutionary algorithm can automatically construct and compress quantum circuits, dramatically reducing circuit depth while keeping high fidelity to a target state. The authors test two variants: one that periodically re-optimizes rotation-gate parameters and one that relies purely on mutation and crossover. On random 4- and 6-qubit circuits, the best from-scratch depth reductions reach 80–90% with fidelities between roughly 0.7 and 0.98, depending on circuit size and starting point. This matters because circuit depth is a direct proxy for noise exposure on current quantum hardware, so automated depth reduction is a practical step toward running useful circuits.","feed_headline":"Hybrid evolution shrinks quantum circuits by up to 90%","feed_subtitle":"Automatic search cuts circuit depth 80–90% at high fidelity-to-target; useful before running on noisy hardware.","key_machinery":"The central object is the circuit solution matrix: rows correspond to qubits and columns to time layers, with each cell holding a gate object, so the whole circuit is a fixed-size grid that mutation and crossover can act on directly. The mechanism is an evolutionary loop combining single-point and uniform-column crossover with a menu of mutations (gate replacement, gate and column swaps, adding or deleting layers, adding CX gates), followed by survivor selection. Fitness is a weighted two-term objective, $f = \\alpha F(U,\\rho) - \\beta \\delta_{\\text{norm}}$, with fidelity $F(U,\\rho)$ to the target state and normalized depth $\\delta_{\\text{norm}}$, using weights $\\alpha = 10$ and $\\beta = 1$. Two subroutines carry the hybrid part: a classical parameter optimizer that adjusts rotation angles on a random 10% of the population every 25 generations to maximize fidelity, and a solution-compaction heuristic that removes identity-only layers and merges rotations without changing the circuit's function.","core_discovery":"The paper's central discovery is that a multi-objective evolutionary algorithm operating on a matrix encoding of quantum circuits can substantially reduce circuit depth while keeping the state close to the target. In the strongest from-scratch result, a random 4-qubit circuit of depth 20 is reproduced with fidelity 0.98354 after an 80% depth reduction by the hybrid EA; the regular EA reaches 82.5% reduction at fidelity 0.9741. Across random circuits, both EAs beat the no-evolution baseline on depth, and the parameter-optimization subroutine generally raises fidelity at the cost of some depth reduction. On 6-qubit circuits, from-scratch depth reductions reach 86–91% with fidelities of 0.71–0.94. When the population is initialized with the target circuit, all EA variants hold fidelity above 0.99 on 4-qubit circuits and still cut depth by 14–43%.","pith_inferences":["The reported 80–90% reductions are on randomly generated circuits whose structure may be unusually compressible; a natural test is to run the same EA on structured benchmark circuits such as QFT, GHZ preparation, or QAOA ansätze, where reductions could be smaller.","Because the fitness weights strongly favor fidelity ($\\alpha=10$, $\\beta=1$), the Pareto frontier between fidelity and depth is not explored; varying the weight ratio could yield even lower-depth circuits at modest fidelity cost, or higher fidelity at less compression.","Depth alone ignores gate count and hardware connectivity, so the practical gain on real devices needs to be checked against transpiled gate counts and error rates; two circuits with the same depth can differ substantially in noise susceptibility."],"forward_implications":["For a fresh random 4-qubit depth-20 circuit, the hybrid EA reduces depth by about 80% while the regular EA reaches 82.5%, both with fidelity above 0.97.","Initializing the population with the target circuit preserves fidelity above 0.99 on 4-qubit circuits up to depth 47, with depth cuts of 14–43%.","From-scratch construction yields larger depth reductions than target-initialized optimization, so the two modes trade fidelity against compression.","Adding parameter optimization improves fidelity in most from-scratch runs but often reduces the depth improvement relative to the pure EA, and it is computationally more expensive.","Turning off the solution-compaction subroutine lowers depth improvement, for example from 80.32% to 77.13% for the regular EA at depth 47 and from 39.89% to 9.57% for the baseline."],"supporting_citations":[{"why":"Supplies the solution-matrix encoding and the evolutionary operations that this EA is directly inspired by.","marker":"[32]"},{"why":"Establishes the multi-objective genetic algorithm approach to quantum circuit discovery that this work builds on and compares against.","marker":"[24]"},{"why":"Provides the quantum-circuit toolkit used to convert solution matrices into executable circuits and to compute fidelity.","marker":"[15]"},{"why":"Supplies the classical parameter optimizer used inside the hybrid parameter-optimization subroutine.","marker":"[22]"},{"why":"Shows prior evolutionary synthesis of quantum circuits, motivating the EA-based construction approach.","marker":"[18]"},{"why":"Applies evolutionary search to quantum architecture design, a closely related approach in the same problem space.","marker":"[34]"}],"fun_headline_variants":["Hybrid EA trims quantum circuits by up to 90%","Evolutionary search cuts quantum depth 90%, keeps fidelity","Quantum circuit depth slashed 90% by hybrid EA","EA hybrid cuts quantum circuits 90% with high fidelity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that four random seeds and the particular randomly generated circuits used are representative of typical circuits, so the reported mean depth reductions and fidelities reflect the method's usual behavior rather than lucky draws.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid EA trims quantum circuits by up to 90%","Evolutionary search cuts quantum depth 90%, keeps fidelity","Quantum circuit depth slashed 90% by hybrid EA","EA hybrid cuts quantum circuits 90% with high fidelity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000742,"raw_usage":{"total_tokens":3276,"prompt_tokens":873,"completion_tokens":2403,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":2334}},"tokens_in":489,"tokens_out":2403,"duration_ms":17386,"temperature":1.0,"reasoning_tokens":2334,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:37:25.754516+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same experiments on a larger set of seeds, say 50 or 100, and on structured benchmark circuits such as QFT, GHZ preparation, or QAOA ansätze; if the mean depth reduction on 4-qubit depth-20 circuits falls well below 80% at comparable fidelity, or if some benchmark circuits show no depth reduction at all, the central claim would be weakened.","supporting_citations":[{"cited_title":"Multi-objective evolutionary algorithms for quantum circuit discovery","cited_arxiv_id":"1812.04458","evidence_quote":"Establishes the multi-objective genetic algorithm approach to quantum circuit discovery that this work builds on and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows prior evolutionary synthesis of quantum circuits, motivating the EA-based construction approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Applies evolutionary search to quantum architecture design, a closely related approach in the same problem space."}],"review_version":1}