{"id":"9326e42b-a6d3-45d5-8d35-d6ff3df37708","arxiv_id":"2505.13695","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A frequency-multiplexed Gaussian boson sampler with over 400 modes can serve as a quantum reservoir computer, with correlation-based features and squeezed light outperforming mean-field and classical-light baselines.","lead":"Researchers used a 400+ mode Gaussian boson sampler as a quantum reservoir computer and classified vowels and handwritten digits. Feeding the classifier measured correlations between light modes, rather than only average photon counts, improved accuracy by more than 20 percentage points on some tasks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Squeezed-vs-coherent advantage hinges on an underspecified simulated coherent reservoir; if it omits loss/detector effects, the central comparison is not established.","rationale":"The paper's central correlation-advantage finding is well supported by direct experimental comparisons within the squeezed-light data. The vulnerable point is the controlled comparison with classical light, exactly where the only equal-bandwidth data point is a simulation. This concern was identified by the reader. I do not see it as fatal because the authors are transparent about the limitation, do not claim a quantum advantage over all classical methods, and the correlation-advantage claim stands independently. I therefore keep the conditional verdict. The requested check is feasible with their released code and data, and it would settle whether the squeezed-light comparison is real or an artifact of an idealized coherent baseline.","tokens_in":104,"tokens_out":6430,"duration_ms":117667,"concrete_test":"Run the released data/code (Zenodo 15288036) to recompute the Fig. 4a/b orange curve using the same experimental noise model as the squeezed-light data: include measured 60% end-to-end transmission, EMCCD quantum efficiency and stochastic gain, per-pixel dark counts, and the actual AFC phase masks; also feed a coherent state with complex spectral amplitudes matched to the squeezed-state mean field rather than only intensities. If the orange curve shifts by more than the reported error bars in either panel, the squeezed-light-over-coherent claim is not supported and should be downgraded to an open question.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's second central claim—that squeezed light consistently gives highest accuracies—is supported against coherent light mainly by the simulated broadband-coherent reservoir (Fig. 4a/b, orange). The experimental coherent source was a CW laser with a different bandwidth and mode structure, so the simulation is the only equal-bandwidth comparison. The paper does not specify how this simulation handles the losses and detector effects that are present in the squeezed-light path: end-to-end transmission is only ~60%, EMCCD quantum efficiency 95%, stochastic EM gain, and per-pixel background. If the simulation instead models an ideal coherent state (shot-noise-limited diagonal covariance, no loss, no detector noise), or if it constructs the coherent state from only the measured mean photon numbers without the correct complex spectral amplitudes and phase relationships that the AFC unitary would imprint, then the orange curve is an idealized lower bound for classical performance rather than a faithful matched baseline. Because the authors themselves note the squeezed-state correlations are not strong enough to exclude a classical source, an artificially weak coherent baseline would directly create the observed superiority without requiring any genuine benefit from squeezing. Thus the claim that squeezed light outperforms equal-bandwidth coherent light is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports an experimental study in which a frequency-multiplexed Gaussian boson sampler with more than 400 modes is used as a reservoir for classification tasks. Classical data are encoded in the spectral phase of the pump of an adiabatic frequency conversion process, and the reservoir output is read out either as the mean photon number per mode or as the full photon-number covariance matrix computed from EMCCD camera frames. The authors evaluate the resulting reservoir computer on synthetic nonlinear tasks, spoken-vowel classification, and MNIST digit classification. Their central claims are that using the covariance matrix instead of the mean-field vector increases classification accuracy, in several cases by more than 20 percentage points, and that a squeezed-light source gives consistently higher (or tied-highest) accuracy than the tested classical light sources under matched total photon budgets. The paper is candid about its main limitations: the classical sources did not have the same bandwidth as the squeezed source, the end-to-end loss is high, and the measured correlations are not strong enough to rule out classical generation.","tokens_in":27488,"tokens_out":5319,"duration_ms":55506,"significance":"If the claims hold, this is a substantial experimental step for photonic quantum reservoir computing: it demonstrates a >400-mode GBS-based reservoir with >100 programmable input dimensions, provides evidence that two-mode correlation features improve linear classification, and compares several classical optical reservoirs under matched photon budgets. The manuscript has notable strengths: the photon-budget matching in Appendix A, the robustness check across four different output classifiers in Fig. 4d, the explicit discussion of loss, classical simulability, and the impossibility of attributing the results to quantumness in Sections IV B and IV F, and the public availability of data and code (DOI in the Data Availability statement). The main weakness is that the controlled equal-bandwidth comparison between squeezed light and coherent light rests on an underspecified simulation, and some plots labeled as classical-reservoir comparisons are actually comparisons between covariance and mean-field outputs of the squeezed-light data.","major_comments":[{"comment":"The simulated broadband coherent-state reservoir (orange curve in Figs. 4a and 4b) is the only equal-bandwidth classical baseline, but the manuscript does not specify the simulation model. It must state how the coherent state is constructed (for example, which measured quantities determine the complex displacement amplitudes and phases), whether the AFC unitary and the approximately 60% end-to-end transmission, the 95% EMCCD quantum efficiency, the stochastic EM gain, and the per-pixel background are included, and how the simulated camera frames are generated. Without this information, the orange curve could be an idealized lossless coherent baseline, and the claim that squeezed light outperforms an equal-bandwidth coherent reservoir would not be established; please provide the model equations or pseudocode and validate the simulator against the experimental coherent-state data shown in the red curves.","section":"Section III B, Fig. 4, and Methods V A"},{"comment":"In several places the 'classical reservoir' is not a reservoir driven by coherent light but the mean-field output computed from the squeezed-light data. The caption of Fig. 10 states that both curves are 'computed from the same measurement datasets,' and Fig. 4c similarly compares the covariance with the mean field. This confounds the choice of output feature (covariance versus mean) with the choice of input state (squeezed versus coherent), and therefore cannot support the abstract's claim that squeezed light outperforms coherent light; either perform the comparison with actual coherent-light datasets or relabel these curves and their discussion as feature-ablation results.","section":"Appendix C, Fig. 10, and Fig. 4c"},{"comment":"The comparison between covariance and mean-field inputs is not symmetric. The covariance input is feature-selected using an ANOVA F-test with the number of retained features k optimized on an 8% validation set, whereas the mean-field vector is always used with all 512 components; moreover, the covariance matrix contains the diagonal elements, which carry mean-field-like information. To support the statement that inter-mode 'correlations' are responsible for the reported gains of greater than 20 percentage points, the authors should compare against diagonal-only and off-diagonal-only covariance inputs and use nested cross-validation for the feature-count selection; otherwise the advantage could be due to feature selection or to a rescaling of the diagonal information rather than to the correlations between modes.","section":"Section III A and Appendix D"},{"comment":"The squeezed-versus-coherent advantage is described as 'consistently' higher in accuracy, while the same footnote states that the gaps are within statistical uncertainty. Since this comparison is load-bearing for the second central claim, the abstract and Section IV should either report explicit confidence intervals or combined statistical tests across tasks, or the wording should be qualified to state that the squeezed-light advantage is suggestive but not statistically established at the reported sample sizes.","section":"Footnote 7 and Section IV B"}],"minor_comments":[{"comment":"The sentence 'in all cases the AFC was pump did not encode an input vector x' contains a grammatical error and should be rewritten.","section":"Section II, setup description"},{"comment":"The text '30, 0000 frames' appears to be a typo; it should presumably read '30,000 frames'.","section":"Section III B"},{"comment":"The color scales and axis units for the covariance maps are not defined in the figure captions; please add colorbars and clarify whether the axes are camera-pixel indices or wavelengths.","section":"Fig. 2 and Fig. 16"},{"comment":"The summation in Eq. (E1) is over m instances, but the notation m is not defined at that point and the Lagrangian formulation in Eq. (E3) also uses m; please make the notation consistent.","section":"Appendix E, Eq. (E1)"},{"comment":"The terms 'mode' and 'camera pixel' are used nearly interchangeably; since the grating maps the AFC output frequencies onto pixels, please state explicitly the pixel-to-mode relation and whether adjacent pixels correspond to distinct optical modes or to an oversampled spectrum.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the manuscript is unusually honest about its limitations, and the main concerns are fixable in revision. The most important issue is the underspecified simulated coherent baseline, since it carries the equal-bandwidth comparison that the abstract relies on; the authors' own Footnote 7 and Section IV B already concede the main weakness, so the revision should align the abstract and the discussion with that caution and provide the missing simulation details. I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'll get straight to it. This is the first experimental reservoir computer built on a frequency-multiplexed GBS with over 400 modes, and that alone makes it worth attention. The authors do two things well: they show, across several tasks, that feeding a linear classifier the measured covariance matrix consistently matches or beats the mean-field vector, sometimes by more than 20 points; and they make a serious attempt to compare squeezed light against three classical sources under matched photon budgets. The data and code are posted, and the paper is unusually honest about what it does not claim.\n\nThe correlation result is solid. The squeezed-versus-coherent comparison is the soft spot. The experimental coherent source was a CW laser with a different bandwidth and mode structure, so the only equal-bandwidth comparison is a simulated broadband coherent reservoir (the orange curve). No one reading the paper can tell from the text whether that simulation includes the loss, detector quantum efficiency, stochastic EM gain, and per-pixel background that the squeezed path actually saw. If it does, the orange curve is a fair classical benchmark and the claim is interesting. If it is an ideal lossless coherent-state simulation, it is a strawman, and the 'squeezed light consistently gave the highest accuracies' claim is not established on the evidence presented. The authors hedge in the discussion, but the abstract states it as a finding. This needs to be fixed in revision, probably by spelling out the simulation model and ideally running the same simulation with realistic noise.\n\nTwo smaller things. Appendix C contains a claim that accuracy scales exponentially with the number of modes. The plotted data plateau well before 400 modes; 'exponential' describes a short segment, not a scaling law, and should be reworded. And the feature-selection protocol uses a validation set to pick k; the paper should state clearly whether those same samples are then reused in training the final classifier, and what that does to the reported accuracies if they leak.\n\nNone of this changes my bottom line. The paper is a solid experimental platform paper for the photonic reservoir computing community, with a reproducible core result about correlations. It deserves a serious referee, and I'd send it to review with a request to pin down the coherent-state simulation and to soften the exponential claim. If the authors can show the simulated coherent baseline includes realistic loss and noise, the squeezed-light result stands; if not, the abstract needs to be dialed back.","headline":"A genuinely useful first demonstration of >400-mode GBS reservoir computing, with correlation features clearly helping, but the squeezed-versus-coherent superiority claim depends on an underspecified simulated baseline that needs to be pinned down before it convinces.","tokens_in":28089,"tokens_out":2412,"would_cite":true,"duration_ms":24166,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A >400-mode Gaussian boson sampler serves as a reservoir computer whose accuracy improves by over 20 percentage points when the output layer is given photon-number correlations between modes rather than only per-mode averages.","keywords":["quantum reservoir computing","Gaussian boson sampling","squeezed light","photon correlations","covariance matrix","machine learning","adiabatic frequency conversion","optical computing"],"falsifier":"Repeat the vowel and moons/blobs classification experiments with a coherent light source whose spectral bandwidth, loss, mode structure, and photon budget exactly match the squeezed-vacuum path; if its covariance-based accuracy reaches or exceeds the squeezed-light accuracy within error bars, the claim that squeezing itself provides the correlation advantage is refuted.","tokens_in":27075,"feed_emoji":"🔬","tokens_out":8921,"duration_ms":72762,"temperature":0.7,"pith_summary":"This paper reports experiments using a frequency-multiplexed Gaussian boson sampler with more than 400 optical modes as the fixed reservoir in quantum reservoir computing. The authors classify classical data—synthetic non-linearly separable tasks, spoken vowels, and MNIST digits—by encoding each sample in the spectral phases of a pump pulse, measuring output photon counts per mode, and training a single classical linear classifier on the measurements. The central result is that feeding the classifier the full photon-number covariance matrix, which includes correlations between modes, gives the same or higher accuracy than feeding it only the per-mode mean photon numbers, with gains greater than 20 percentage points in several cases. They also compare squeezed light against coherent, thermal, and supercontinuum light under matched photon budgets and find squeezed light consistently at least ties the best classical accuracy. The work positions GBS as a practical hardware platform for quantum reservoir computing at large system sizes.","feed_headline":"Photon correlations, not counts, boost quantum reservoir accuracy","feed_subtitle":"A 400-mode Gaussian boson sampler classifies digits and vowels better when its output layer sees mode-mode correlations.","key_machinery":"The workhorse is the photon-number covariance matrix of the measured output modes, $\\Sigma_{ij} = \\langle \\hat n_i \\hat n_j\\rangle - \\langle \\hat n_i\\rangle\\langle \\hat n_j\\rangle$, built from single-shot EMCCD camera frames and flattened into a feature vector for a linear classifier. The reservoir itself is an adiabatic frequency-conversion (AFC) crystal acting as a programmable frequency-domain beamsplitter unitary: classical data $x$ are encoded in the spectral phases $\\phi(\\lambda)$ of the pump pulse via a spatial light modulator, and the weak input optical state (squeezed vacuum from an OPA, or coherent, thermal, or supercontinuum light) is transformed by this unitary before frequency-resolved detection. The covariance's off-diagonal elements capture inter-mode correlations that the mean-field vector discards; the paper argues that these correlations are what let the linear classifier separate classes that are not linearly separable in the raw features.","core_discovery":"The paper claims that a Gaussian boson sampler used as a reservoir computer produces features whose correlations are computationally useful: across every benchmark task, providing the trained linear output layer with the photon-number covariance matrix $\\Sigma$ (with elements $\\Sigma_{ij} = \\langle \\hat{n}_i \\hat{n}_j\\rangle - \\langle \\hat{n}_i\\rangle \\langle \\hat{n}_j\\rangle$) rather than only the mean-field vector $\\mu$ yields equal or higher classification accuracy. For the squeezed-light reservoir, this advantage exceeded 20 percentage points in several configurations, and the accuracy gap grows with task complexity, number of modes read out, and photon budget. The paper further claims that squeezed light used as the reservoir input consistently achieves the highest or tied-highest accuracies among squeezed, coherent, thermal, and supercontinuum light when total detected photon numbers are matched, with the caveat that only a simulated same-bandwidth coherent source provides an equal-bandwidth comparison.","pith_inferences":["Beyond the paper: if correlation-based features are as informative as these experiments suggest, other multimode squeezed-light platforms could be repurposed as reservoirs without changing the training protocol, making the AFC setup one instance of a broader class of Gaussian reservoir hardware.","Beyond the paper: the quadratic growth of the covariance matrix with mode count predicts that correlation readouts should become increasingly valuable relative to mean-field readouts as the number of modes grows; a direct test would compare accuracy gaps at, say, 64, 128, 256, and 400 modes.","Beyond the paper: the authors note their system could handle time-series inputs; a testable extension is to run temporal tasks (e.g., waveform forecasting) to see whether covariance readouts also outperform mean-field readouts in the recurrent setting.","Beyond the paper: the loss caveat suggests a crisp experiment—inserting a lower-loss detection stage and checking whether the correlation-driven accuracy gap widens once the measured correlations are verifiably non-classical."],"forward_implications":["Access to correlations in the reservoir output should be built into QRC feature design; mean-field-only readouts may leave substantial accuracy on the table.","Larger numbers of readout modes and higher photon budgets both increase accuracy when correlations are used, so scaling GBS reservoirs in these directions should improve performance on harder tasks.","Squeezed light is worth retaining as a resource even though the current system's loss keeps the correlations from being unambiguously quantum; a lower-loss version could decide whether quantum correlations genuinely cause the advantage.","The same GBS apparatus, operated with quantum states rather than classical data as inputs, could be used to classify quantum states of light, a task where QRC might deliver a true quantum advantage.","On the MNIST benchmark, the achieved accuracy is comparable to a recently reported neutral-atom quantum reservoir computer, indicating that photonic GBS is a competitive large-scale QRC platform."],"supporting_citations":[{"why":"supplies the experimental apparatus: the highly multimode squeezed-light source and the programmable adiabatic-frequency-conversion unitary used as the reservoir.","marker":"[47]"},{"why":"the theoretical prediction that access to correlations can enhance the power of quantum reservoir computers, which the paper's correlation results support.","marker":"[24]"},{"why":"the theoretical framework for Gaussian-state reservoir computing that the GBS-based reservoir instantiates.","marker":"[44]"},{"why":"the foundational quantum reservoir computing proposal motivating the use of an untrained quantum system as a reservoir.","marker":"[22]"},{"why":"the standard reservoir-computing training protocol (linear readout trained on reservoir features) that the paper follows.","marker":"[48]"},{"why":"the comparable MNIST quantum-reservoir result on an analog quantum computer that the paper uses as a benchmark.","marker":"[19]"},{"why":"supplies the spoken-vowels dataset used for the main real-world classification benchmark.","marker":"[60]"},{"why":"supplies the MNIST handwritten-digit dataset used for the high-dimensional benchmark.","marker":"[61]"},{"why":"supports the acknowledged limitation that the lossy GBS with mean-field and covariance measurements is likely classically simulable.","marker":"[13]"}],"fun_headline_variants":["Photon correlations boost quantum reservoir accuracy","Gaussian boson sampler thrives on mode correlations","Squeezed light beats classical in quantum reservoir tasks","Correlations, not counts, key for quantum reservoir computing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison between squeezed and coherent light hinges on the simulated broadband coherent reservoir faithfully representing a classical reservoir with the same mode structure, loss, and encoding as the squeezed-light experiment; if that simulation omits loss, mode overlap, or detector effects present in the real squeezed-light path, the claim that squeezing itself helps is not established.","fun_headline_variants_meta":{"raw":{"variants":["Photon correlations boost quantum reservoir accuracy","Gaussian boson sampler thrives on mode correlations","Squeezed light beats classical in quantum reservoir tasks","Correlations, not counts, key for quantum reservoir computing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00026,"raw_usage":{"total_tokens":1607,"prompt_tokens":977,"completion_tokens":630,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":570}},"tokens_in":593,"tokens_out":630,"duration_ms":5103,"temperature":1.0,"reasoning_tokens":570,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:11:26.059691+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the vowel and moons/blobs classification experiments with a coherent light source whose spectral bandwidth, loss, mode structure, and photon budget exactly match the squeezed-vacuum path; if its covariance-based accuracy reaches or exceeds the squeezed-light accuracy within error bars, the claim that squeezing itself provides the correlation advantage is refuted.","supporting_citations":[{"cited_title":"Nokkala, R","cited_arxiv_id":null,"evidence_quote":"supplies the experimental apparatus: the highly multimode squeezed-light source and the programmable adiabatic-frequency-conversion unitary used as the reservoir."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"the theoretical prediction that access to correlations can enhance the power of quantum reservoir computers, which the paper's correlation results support."},{"cited_title":"Killoran, T","cited_arxiv_id":null,"evidence_quote":"the theoretical framework for Gaussian-state reservoir computing that the GBS-based reservoir instantiates."},{"cited_title":"Mujal, R","cited_arxiv_id":null,"evidence_quote":"the standard reservoir-computing training protocol (linear readout trained on reservoir features) that the paper follows."},{"cited_title":"Rosskopf, T","cited_arxiv_id":null,"evidence_quote":"supplies the spoken-vowels dataset used for the main real-world classification benchmark."}],"review_version":1}