{"id":"68049776-5206-440b-bdf8-a1c88d558e9f","arxiv_id":"2502.10242","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An experimental continuous-variable quantum compiler learns an optical phase with two-mode squeezed light, and increasing the squeezing sharpens the cost landscape, improving precision and training speed.","lead":"Researchers demonstrated a quantum light-based compiler that learns an unknown optical phase by squeezing two correlated light beams and tuning a control phase until it matches the target. In their setup, higher squeezing gave up to 5.4 times better phase precision and 3.6 times faster convergence.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed genuine quantum advantage is not established: no equal-energy coherent-state baseline is provided, and the Eq. (3) cost function is phase-insensitive for coherent states, so the 5.4x/3.6x factors cannot be attributed to squeezing correlations without a control.","rationale":"The strongest part of the paper is the direct measurement of the cost landscape and its narrowing with squeezing (Fig. 3a), which is independent of the fitted model and shows a real experimental effect. The authors also demonstrate convergence for arbitrary target and control phases (Appendix E) and validate their data processing with a complementary method. However, the interpretative leap to 'genuine quantum advantage' in the Discussion requires showing that the improvements are due to quantum correlations rather than to the increased energy and improved signal-to-noise ratio that accompany higher r. The paper's own cost function (Eq. 3) cannot be evaluated for coherent states, so the theoretical analysis in Eqs. (8-10) only compares within the squeezed-state manifold. A control experiment with a coherent-state resource using a first-moment cost function would settle this. The precision headline also rests on N=5 runs without confidence intervals; Appendix B partially addresses this with N=15, but the reported ratio still lacks uncertainty. These issues are fixable and do not invalidate the raw measurements, so the reader's CONDITIONAL verdict stands.","tokens_in":24784,"tokens_out":7423,"duration_ms":75181,"concrete_test":"Repeat the QCA with a coherent-state resource at matched total photon number and identical per-iteration acquisition time, using a phase-sensitive cost function based on the first moment of the homodyne distribution (e.g., C = |⟨q_-⟩|). If this classical baseline yields comparable or better precision-per-photon and time-to-solution than the squeezed-resource QCA, the quantum-advantage claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that quantum correlations, not total photon number, drive the observed improvements (Discussion, Section IV) is unsupported because no equal-energy classical baseline is reported. The cost function defined in Eq. (3) is constructed from the quadrature-noise marginal of a TMSS; for a coherent state this marginal is phase-independent, so the paper's own cost cannot be used to compare against a classical resource. The theoretical interpolating analysis (Eqs. 8-10) lives on the seeded-TMSS manifold and only shows that within that manifold, squeezing improves the cost landscape curvature and precision scaling; it does not compare to an optimized coherent-state phase estimator with the same total photon budget and measurement time. Consequently, the reported 5.4x precision increase (Table I, from N=5 runs without error bars) and 3.6x iteration-speedup could be caused by increased photon number, increased local-oscillator power, or stronger signal-to-noise ratio rather than by two-mode squeezing correlations. Without a control experiment or a resource-accounted analysis, the 'genuine quantum advantage' claim does not follow from the data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports an experimental continuous-variable (CV) quantum compiler that learns a single optical phase gate via gradient descent. The resource is a two-mode squeezed state (TMSS) generated by four-wave mixing in a truncated SU(1,1) interferometer. The authors derive a cost function from the amplitude-difference marginal of the Wigner function, fit their measured cost curves to a noisy-TMSS model with two free parameters, and show that increasing the effective squeezing parameter r narrows the cost landscape, speeds convergence (3.6-fold reduction in iterations to convergence), and improves the reported phase precision (5.4-fold between r=0.18 and r=0.74). They also present a theoretical analysis of the cost landscape and claim that the observed enhancements constitute a 'genuine quantum advantage' driven by quantum correlations rather than total photon number. The paper includes detailed appendices on data processing, parameter estimation, and robustness checks.","tokens_in":25047,"tokens_out":4614,"duration_ms":51750,"significance":"If supported, this would be a valuable first experimental demonstration of a squeezing-based CV variational quantum compiler, with a practical cost function, direct comparison to theoretical predictions, and a tunable landscape that connects to barren-plateau avoidance in the CV setting. The theoretical cost-function derivation is explicit, and Eq. (5) provides a parameter-free prediction of the lossless precision scaling given the squeezing parameter, which is a notable strength. The experimental methods are detailed, including noise modeling, AIC-based model comparison, and robustness tests over target/control phases. However, the central quantitative claims—the 5.4-fold and 3.6-fold factors and, especially, the asserted 'genuine quantum advantage'—are not fully supported by the data as presented, so the significance of the paper currently rests on the more modest demonstration of a controllable CV compiler.","major_comments":[{"comment":"The claim of a 'genuine quantum advantage' over an equal-energy classical (coherent-state) resource is not supported by the data. No equal-energy coherent-state control experiment is reported, and the cost function defined in Eq. (3) is phase-insensitive for coherent states because the quadrature-noise marginal of a coherent state has constant variance independent of Δφ. The interpolating analysis in Eqs. (8)–(10) lives on the seeded-TMSS manifold and does not compare against an optimized coherent-state phase-estimation strategy with the same total photon number and measurement time. The observed 5.4-fold precision increase and 3.6-fold speedup could in principle arise from increased photon number, higher local-oscillator power, or improved signal-to-noise rather than from two-mode squeezing correlations. To substantiate the 'quantum correlations, not merely total photon number' statement, the authors should either add a resource-accounted comparison with a classical baseline using a phase-sensitive readout, or explicitly temper the claim to a demonstration of enhanced precision and speed with squeezed resources without asserting genuine quantum advantage.","section":"Section IV"},{"comment":"The 5.4-fold precision factor rests on N=5 runs per squeezing value with no uncertainty reported on σΔφ, and the N=15 validation run in Appendix B yields a smaller factor of 4.15 with a modified learning rate. The paper does not provide error bars, confidence intervals, or a statistical test that the precision ratio differs from what could be obtained from noisy data. Please report the uncertainty on the σΔφ estimates (e.g., bootstrap or chi-square intervals) or present the factor as a qualitative trend rather than a precise quantitative claim. In addition, the claim that the measurements are 'unbiased' is not persuasive given the relatively large phase-difference means in Table I (e.g., 90 mrad for r=0.18) and especially in Table II (485 mrad for r=0.18), where the authors attribute the offset to the learning rate; this should be discussed quantitatively.","section":"Section III, Table I, and Appendix B"},{"comment":"The inferred squeezing parameters r=0.18, 0.35, and 0.74, and hence the theoretical cost curves and the quoted precision scaling, depend on a model that the authors themselves state neglects distributed gain and loss in the four-wave mixing medium and underestimates the effective input noise Nin. The fitting procedure constrains r through a relation r(Nin, ε') that is derived from this simplified model, so the reported ratios and the 'validated' scaling could shift under a different noise model. Please provide a sensitivity analysis showing how the extracted r values and the resulting precision/speed-up factors change under the alternative noisy-homodyne model (or a model that incorporates distributed loss), rather than only reporting the AIC difference.","section":"Appendix D"}],"minor_comments":[{"comment":"The text refers to the main experimental schematic as Fig. 4(a) and Fig. 4(b), but the schematic with panels (a) and (b) appears to be Fig. 1; the later Fig. 4 is the data-processing validation diagram. Please correct the cross-references.","section":"Section II A and II C"},{"comment":"There is a typo: 'homoydne detection' should be 'homodyne detection'.","section":"Section II A"},{"comment":"The sentence 'expressed as, expressed as' contains a duplicated phrase.","section":"Appendix D"},{"comment":"The text 'over the thee runs' should read 'over the three runs'.","section":"Appendix E"},{"comment":"The quantity σΔσ in the paragraph describing Table II should be σΔφ.","section":"Appendix B"},{"comment":"The figure title says 'complimentary data processing procedure'; 'complimentary' should be 'complementary'.","section":"Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper reports a well-executed experiment with careful noise characterization and validation, and it fits well within the scope of a quantum information journal. The main obstacle to acceptance is the over-reaching 'genuine quantum advantage' claim in Section IV, which is not supported by an equal-energy coherent-state baseline or a resource-accounted comparison. A revision that removes or substantially qualifies this claim and adds sensitivity analysis for the model-dependent r values would make the contribution solid. I recommend major revision rather than rejection because the core experimental demonstration appears sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a real experimental implementation of a CV quantum compiler using two-mode squeezed light, the first I know of, and the cost-landscape narrowing with squeezing is clearly shown. But the paper's strongest claim—'genuine quantum advantage'—is not supported by the data, because there is no equal-energy coherent-state control. The reported 5.4x precision factor is also fragile: N=5 runs, no error bars, and the N=15 validation gives 4.15 under a different learning rate.\n\nThe experimental work itself is solid. The authors implement the TMSS-based cost function from Volkoff et al., show that the measured cost curves track the theory across three squeezing levels, and demonstrate that higher squeezing gives faster convergence and lower phase variance. The seeded-TMSS precision analysis around Eqs. (8)-(10) is a nice addition, showing interpolation between sub-SQL and Heisenberg-like scaling within that manifold. Credit is due for transparency: Appendix D admits that the model ignores distributed gain and loss, and the paper explicitly notes that the Eq. (3) cost function is meaningless for pure coherent states.\n\nThe soft spots are real but not fatal to the core demonstration. First, the 'genuine quantum advantage' claim in the Discussion outruns the evidence. Since increasing r also increases total photon number, the observed improvements could in principle come from energy rather than correlations. The paper needs a coherent-state baseline with the same photon budget, or a resource-accounted comparison, before claiming the correlations are doing the work. The fact that Eq. (3) can't be used for coherent states means the authors would need a different readout for the control, but that's an experimental challenge, not a get-out-of-jail card. Second, the headline precision factor is a point estimate from five runs with no uncertainty; the N=15 study gives 4.15 with a different learning rate, so the exact number should not be taken as precise. Third, the squeezing parameters come from a fitted noisy-TMSS model whose limitations are acknowledged; the fits look good and are constrained by independent squeezing measurements, so this is a moderate concern, not a fatal one.\n\nOverall: this is a solid proof-of-principle experiment that deserves peer review. It should be published after the authors either pull the quantum advantage claim back to what is actually shown (tunable cost landscape, faster convergence, improved precision with squeezing) or add a proper classical baseline. I'd send it to a serious referee.","headline":"Real first demonstration of a CV quantum compiler, but the 'genuine quantum advantage' claim lacks an equal-energy coherent-state control and the headline precision factor rests on N=5.","tokens_in":25567,"tokens_out":3369,"would_cite":false,"duration_ms":34634,"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":"This paper reports an experimental continuous-variable quantum compiler that learns the phase of an optical gate using a two-mode squeezed state, with a 5.4-fold precision gain and a 3.6-fold faster convergence.","keywords":["continuous-variable quantum computing","quantum compiling","two-mode squeezed state","variational quantum algorithm","optical phase estimation","homodyne detection","four-wave mixing","barren plateaus"],"falsifier":"Lock the apparatus at each detuning and measure the amplitude-difference squeezing directly at $\\Delta\\phi=0$ with an independent characterization, then check whether the observed phase standard deviation and time-to-solution follow the $f(N)$-based scalings with that independently measured $r$; separately, run the compiler with a coherent-state probe of equal total photon number, because if its precision and speed match the squeezed-resource results, the claimed quantum advantage is falsified.","tokens_in":24621,"feed_emoji":"🔬","tokens_out":7379,"duration_ms":74737,"temperature":0.7,"pith_summary":"This paper reports the first experimental continuous-variable quantum compiler: a circuit that learns an unknown optical phase shift by iteratively adjusting a control phase until its effect matches the target. The resource is a two-mode squeezed state of light, whose quantum correlations let the experiment tune the steepness of the cost landscape by changing the squeezing level. At the highest squeezing studied, the compiled phase estimate was 5.4 times more precise and reached convergence 3.6 times faster than at the lowest squeezing. The paper argues this improvement comes from the quantum correlations themselves, not from sending more photons, and that the same mechanism pushes phase precision toward Heisenberg scaling.","feed_headline":"Squeezed-light compiler learns optical phases 5.4 times more precisely","feed_subtitle":"Two-mode squeezed light narrows the cost landscape, cutting time-to-solution 3.6-fold in a variational quantum compiler.","key_machinery":"The load-bearing object is the cost function $C(r,\\varepsilon,\\varepsilon',N'_B,\\Delta\\phi) = -\\rho(0,r,\\varepsilon,\\varepsilon',N'_B,\\Delta\\phi) = -\\sqrt{2/(\\pi(\\mathrm{tr}A+2A_{1,2}))}$, obtained by projecting the Wigner function of the noisy two-mode squeezed state onto the amplitude-difference quadrature $(q_1-q_2)/\\sqrt{2}$ and taking its peak value at $X_-=0$; $A$ is the $2\\times 2$ matrix in Eq. (4) depending on squeezing $r$, transmission efficiencies $\\varepsilon,\\varepsilon'$, seed excess noise $N'_B$, and phase difference $\\Delta\\phi$. This cost function is evaluated from homodyne measurement statistics and minimized by gradient descent on $\\phi_c$. The squeezing parameter $r$ acts as a landscape dial: it controls the curvature $f(N)$ near the minimum, which enters the quadratic expansion $C \\approx -1 + f(N)\\Delta\\phi^2/2$ and therefore sets both the achievable phase precision and the gradient magnitude available for training.","core_discovery":"Using a two-mode squeezed state produced by four-wave mixing in a rubidium vapor cell, the authors implement a variational quantum compilation algorithm that learns the phase $\\phi_0$ of a target optical phase gate. A control phase $\\phi_c$ is adjusted by gradient descent to minimize a cost function defined as the negative peak value of the homodyne-measured marginal distribution of the amplitude-difference quadrature; minimizing this cost drives $\\phi_c$ to $\\phi_0$. The central discovery is that the squeezing parameter $r$ tunes the shape of this cost function: increasing $r$ narrows the minimum and steepens its slopes, which simultaneously raises the precision with which $\\phi_0$ can be estimated and shortens the time-to-solution. Quantitatively, the measured phase standard deviation falls from 513.8 mrad at $r=0.18$ to 96 mrad at $r=0.74$ (a 5.4-fold precision gain), while convergence time falls from 470 to 130 iterations (a 3.6-fold speedup). In the ideal limit the cost function's curvature grows with $f(N)=\\sqrt{N(N+1)}(2\\sqrt{N(N+1)}+(2N+1))$, so the phase error scales as $\\delta/N^2$ with the squeezed photon number $N$, which the paper presents as evidence for Heisenberg-like scaling and for a genuine quantum advantage.","pith_inferences":["Beyond the paper, a direct head-to-head test against an equal-energy coherent-state probe, rather than only the theoretical comparison in Eq. (8), would put the 'quantum correlations, not photon number' claim on firmer experimental footing.","Beyond the paper, the proposed low-to-high squeezing schedule could be tested by comparing random initialization with the adaptive schedule at the same final squeezing, checking whether barren-plateau avoidance actually delivers the projected precision gains.","Beyond the paper, the cost-function readout relies on one Gaussian marginal; extending it to multi-parameter Gaussian unitaries such as beam splitters would require tracking more than one marginal, and the paper's Appendix C notes that the simple moment-based cost is not universal.","Beyond the paper, if the single-excess-noise model underestimates distributed gain and loss in the medium, the inferred $r$ values may be systematically biased, and a more detailed model would likely change the quantitative precision scaling even if the qualitative trend survives."],"forward_implications":["At higher squeezing, the same compiler yields more precise phase estimates while converging in fewer iterations, so squeezing is a tunable resource rather than a fixed noise cost.","Because the cost curvature grows with $f(N)$, the phase error scales as $\\delta/N^2$ in the ideal lossless case, approaching the Heisenberg limit set by squeezed-light interferometry.","The adaptive strategy of starting at low squeezing and increasing it during training can avoid the barren-plateau regime at high squeezing, making the compiler trainable at precision that would otherwise be inaccessible.","The same cost-function construction generalizes to learning other Gaussian unitary parameters from homodyne data, since Gaussian states are fully characterized by first and second quadrature moments.","The demonstrated quantum advantage, if it holds, means continuous-variable variational compilers can outperform equal-energy classical or coherent-state counterparts on phase learning."],"supporting_citations":[{"why":"Supplies the theoretical continuous-variable quantum-compilation proposal that a two-mode squeezed resource enables learning with a single input-output pair and a squeezing-tunable cost landscape.","marker":"[26]"},{"why":"Establishes squeezed-light phase precision in a Mach-Zehnder interferometer, the baseline against which the Heisenberg scaling of the cost-function readout is compared.","marker":"[47]"},{"why":"Gives the four-wave-mixing scheme in rubidium vapor used to generate the two-mode squeezed resource.","marker":"[59]"},{"why":"Provides the truncated SU(1,1) interferometer configuration on which the experimental setup is based.","marker":"[61]"},{"why":"Argues that entanglement in continuous-variable resources confers a learning advantage, supporting the paper's claim that quantum correlations rather than photon number drive the performance.","marker":"[44]"},{"why":"Supplies the energy-dependent barren-plateau result for bosonic variational circuits used to explain the training failures at high squeezing.","marker":"[50]"}],"fun_headline_variants":["Squeezing boosts phase precision 5.4x, speeds learning 3.6x","Learn optical phases 5.4x sharper, 3.6x faster","CV quantum compiler: squeezing gives 5.4x precision, 3.6x speed","Two-mode squeezing sharpens phase learning 5.4x, cuts time 3.6x","Quantum compiler uses squeezed light for 5.4x precision, 3.6x speedup"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole quantitative story depends on modeling the four-wave-mixing resource as a two-mode squeezed state with one excess-noise parameter and noiseless attenuation; the fit of this model to the homodyne data supplies the $r$ values, the convergence thresholds, and the cost curves, and the paper's own Appendix D notes the model ignores distributed gain and loss in the medium, which would increase the effective input noise.","fun_headline_variants_meta":{"raw":{"variants":["Squeezing boosts phase precision 5.4x, speeds learning 3.6x","Learn optical phases 5.4x sharper, 3.6x faster","CV quantum compiler: squeezing gives 5.4x precision, 3.6x speed","Two-mode squeezing sharpens phase learning 5.4x, cuts time 3.6x","Quantum compiler uses squeezed light for 5.4x precision, 3.6x speedup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000631,"raw_usage":{"total_tokens":2948,"prompt_tokens":1012,"completion_tokens":1936,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":1816}},"tokens_in":628,"tokens_out":1936,"duration_ms":16160,"temperature":1.0,"reasoning_tokens":1816,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T18:49:02.926817+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Lock the apparatus at each detuning and measure the amplitude-difference squeezing directly at $\\Delta\\phi=0$ with an independent characterization, then check whether the observed phase standard deviation and time-to-solution follow the $f(N)$-based scalings with that independently measured $r$; separately, run the compiler with a coherent-state probe of equal total photon number, because if its precision and speed match the squeezed-resource results, the claimed quantum advantage is falsified.","supporting_citations":[{"cited_title":"Encoding a qubit in an oscillator,","cited_arxiv_id":null,"evidence_quote":"Establishes squeezed-light phase precision in a Mach-Zehnder interferometer, the baseline against which the Heisenberg scaling of the cost-function readout is compared."},{"cited_title":"Efficient trainability of linear optical mod- ules in quantum optical neural networks,","cited_arxiv_id":null,"evidence_quote":"Gives the four-wave-mixing scheme in rubidium vapor used to generate the two-mode squeezed resource."},{"cited_title":"Distributed quantum sens- ing in a continuous-variable entangled network,","cited_arxiv_id":null,"evidence_quote":"Argues that entanglement in continuous-variable resources confers a learning advantage, supporting the paper's claim that quantum correlations rather than photon number drive the performance."},{"cited_title":"Nonclassical Properties and Quantum Re- sources of Hierarchical Photonic Superposition States,","cited_arxiv_id":null,"evidence_quote":"Supplies the energy-dependent barren-plateau result for bosonic variational circuits used to explain the training failures at high squeezing."}],"review_version":1}