{"id":"7152401e-1c19-4771-b2dd-d45dcaae7fad","arxiv_id":"2506.21537","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"ResQ encodes classification inputs into the Hamiltonian pulses of an analog Rydberg quantum computer and trains it as a 'residual network,' reporting accuracy gains over classical baselines that may stem from weak baseline setup.","lead":"This paper introduces ResQ, a way to train an analog Rydberg atom quantum computer to do image and health classification by shaping laser pulses that control the atoms. The authors claim ResQ outperforms similarly sized classical neural networks, but the comparison appears to give the classical models only the same few features and a very short training run.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that RESQ implements residual neural networks is unsupported: Eq.","rationale":"The reader's weakest assumption, unfair classical baselines, is valid and important. However, it is not the most fundamental problem with the central claim. Even if the baselines were trained optimally and RESQ still won, the paper would not have shown what it claims unless the construction actually contains residual connections. Section 2.2 defines residual networks by the ODE dx/dt = F(x, theta), but Section 4 never demonstrates that the Schrodinger evolution reduces to this form, and no equation introduces the residual block x + F(x). H(t) in Eq. (1) is a function of encoded inputs and parameters, independent of |psi(t)>, so the quantum channel is a linear, unitary, input-dependent map; the output nonlinearity comes only from measurement probabilities. This makes RESQ a valid analog variational classifier but not a residual neural network in the sense used in the paper. The open-sourced code and real-hardware runs are useful engineering evidence, but they validate a classifier, not the residual-network claim. I therefore agree with the REJECT verdict, but for a more fundamental reason than the reader's stated weakest assumption.","tokens_in":13975,"tokens_out":7113,"duration_ms":87508,"concrete_test":"Trotterize Eq. (1) over one pulse interval and write the discrete state update as |psi_{l+1}> = U_l|psi_l>; check whether this can be algebraically rearranged into x_{l+1} = x_l + F(x_l, theta_l) with a nonlinear F. If the update is only linear in the state and no additive residual term appears, then the paper's claim of implementing residual neural networks fails, and the headline should be re-scoped to 'analog quantum classifier parameterized by Rydberg pulses.'","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim has two load-bearing parts: (i) RESQ is a residual neural network, and (ii) it outperforms similarly sized classical models. Part (i) is the more fundamental and is not established. Section 2.2 defines ResNets via dx/dt = F(x, theta) and residual blocks x + F(x). Section 4.1 instead evolves the quantum state by i*hbar*d|psi>/dt = H(t)|psi>, where H(t) in Eq. (1) depends on the input features omega_i and trainable parameters, but not on the instantaneous state |psi(t)>. Consequently, the Trotterized update is |psi(t+dt)> ≈ (I - iH(t)dt)|psi(t)>, a linear state update with a data-dependent generator; the only nonlinearity in the output comes from the final Born-rule measurement. No equation in Section 4 defines a nonlinear residual function F, a skip connection, or a state variable x that plays the role of the data representation. Calling this an implementation of ResNets is an analogy, not a derivation. Part (ii) is also weakened: the classical baselines receive the same 5 PCA features, are trained for only 75 full-batch iterations, and are not hyperparameter-tuned; near-chance baselines such as 51% on MNIST 0/1 indicate undertraining, so the reported 56%, 57%, and 36% improvements are not reliable evidence of a quantum advantage.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces RESQ, a framework that encodes PCA-reduced input features and trainable parameters into the time-dependent pulse sequences of an analog Rydberg atom quantum computer. The dynamics are governed by the Schrödinger equation with a globally and locally addressed Hamiltonian (Eqs. (1)-(2)), and classification is performed by measuring the average excited-state probability. The authors claim that this constitutes the first implementation of residual neural networks (ResNets) and neural ODEs on analog Rydberg hardware, and that RESQ outperforms similarly sized classical feedforward networks, ResNets, and neural ODE classifiers by 56%, 57%, and 36%, respectively, while remaining robust to hardware noise. Evaluation is performed on MNIST, FashionMNIST, and the Pima Indians Diabetes dataset using classical simulation, with a small set of inference runs on the QuEra Aquila device.","tokens_in":14240,"tokens_out":3187,"duration_ms":36173,"significance":"If the central claims were valid, the paper would represent a notable advance in analog quantum machine learning, combining continuous Hamiltonian evolution with a residual-network-style architecture and demonstrating a practical advantage over classical models. The paper also has concrete strengths: the code and data are open-sourced, the authors use a real 256-qubit Rydberg device for inference, and they explore several lattice geometries and spacings. However, the significance is substantially weakened by two load-bearing problems: (i) the residual-network interpretation is not formally justified, since the Hamiltonian evolution is linear and state-independent, and (ii) the classical baselines are handicapped by receiving the same 5 PCA features and only 75 training iterations, making the reported improvements unsurprising. The hardware robustness claim rests on only 8 samples per task, which cannot support the stated conclusion. The result, as presented, does not establish a quantum advantage or a genuine implementation of residual networks.","major_comments":[{"comment":"The central architectural claim is that RESQ implements residual neural networks, but the dynamics in Eq. (1) do not contain a residual function F(x, θ) of the form defined in §2.2. The Hamiltonian H(t) depends on the input features ω_i and on trainable parameters, but not on the instantaneous quantum state |ψ(t)⟩. Consequently, the Trotterized update |ψ(t+dt)⟩ ≈ (I - iH(t)dt)|ψ(t)⟩ is a linear, state-independent map; the only nonlinearity in the entire model is the final Born-rule measurement. Calling this an implementation of ResNets or neural ODEs is an analogy, not a derivation. This undermines the abstract's primary claim of being 'the first framework to implement residual neural networks on analog Rydberg atom quantum computers.'","section":"§4.1, Eq. (1)"},{"comment":"The performance comparison is not fair. The classical baselines C-NN, C-ResNet, and C-NODE receive the same 5 PCA features as RESQ and are trained for only 75 full-batch iterations with no minibatching and no hyperparameter tuning. Under these conditions, near-chance accuracies such as 51% on MNIST 0/1 (Figure 6) indicate severe undertraining, so the reported improvements of 56%, 57%, and 36% in Section 6.1 are artifacts of weak baselines rather than evidence of a quantum advantage. A meaningful comparison would allow classical models to use the raw data or a reasonable feature representation, and would use standard training budgets with hyperparameter selection.","section":"§5, 'Comparative Classical Techniques'"},{"comment":"The robustness claim on real hardware is based on only 8 samples per classification task. With n=8, the observation that some points near the decision boundary flip under noise is anecdotal and cannot support the conclusion that 'RESQ has excellent resilience to noise.' The simulated-noise analysis reportedly shows accuracy within 1% of ideal, but no error bars, confidence intervals, or detailed statistics are provided. This is insufficient support for the paper's robustness claim.","section":"§6.4, 'Real Hardware Executions'"},{"comment":"The paper does not ablate the choice of 5 PCA features or compare against classical models using more features or raw pixel inputs. Since the classical baselines are restricted to the same 5 features, the comparison conflates the feature representation with the model architecture. Without such ablations, the statement that RESQ 'outperforms similarly sized classical models' is not meaningful, because the classical models are artificially limited in their input information.","section":"§4.4 and §6.1"}],"minor_comments":[{"comment":"The name 'RESQ' is typeset inconsistently as 'R ESQ' in several places; please use a consistent notation.","section":"Throughout"},{"comment":"The caption says 'Pluses used by RESQ'; this should read 'Pulses used by RESQ.'","section":"Figure 5 caption"},{"comment":"The statement that Rydberg atom systems 'are currently the only hardware that supports continuous-time Hamiltonian evolution and local/global control' is too strong and unsubstantiated; other quantum platforms (e.g., trapped ions) also offer continuous control, and the claim should be qualified or removed.","section":"§1"},{"comment":"The training hyperparameters are underspecified. The paper mentions Adam and 20 gradient samples but does not report learning rate, number of gradient steps per iteration, or initialization details beyond 'all parameters initialized to 1.0,' which would be needed to reproduce the results.","section":"§5, 'Software and Simulation Setup'"}],"recommendation":"reject","confidential_remarks":"The manuscript is an arXiv preprint that also claims publication at ICCV 2025. My assessment is based solely on the scientific content. The central claims are not supported: the residual-network interpretation is an analogy without a formal derivation, and the classical baselines are so weak that the reported improvements are not informative. The paper does have useful engineering contributions, such as open-source code and real-hardware experiments, but these do not compensate for the load-bearing flaws. If the authors were to substantially reframe the work as a variational analog quantum classifier (without the ResNet claim) and conduct fair baselines with proper hyperparameter tuning, a resubmission might be considered, but the current form is not acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First: the encoding scheme is a real contribution. RESQ parameterizes pulse amplitudes as piecewise-linear functions θ_j ω_i + θ_{j+1}, uses local detuning couplings to embed features, and demonstrates this on QuEra Aquila with code and data available. The linear scaling of input features with qubit count is a nice property, and the authors deserve credit for testing on real hardware.\n\nThe problem is that the two headline claims are not supported. The paper nowhere derives a residual-network mapping. The Schrödinger evolution is linear in the state; the Trotterized update is |ψ(t+dt)⟩ ≈ (I − iH(t)dt)|ψ(t)⟩, with no nonlinear residual function F(x), no skip connection, and no state variable playing the role of a data representation. The only nonlinearity comes from the final Born-rule measurement. Calling this a quantum neural ODE is an analogy, not a derivation, and the residual claim is load-bearing for the paper's novelty.\n\nThe performance advantage is also an artifact of the baseline setup. Classical models receive the same 5 PCA features and only 75 full-batch iterations, with no hyperparameter tuning. Near-chance scores like 51% on MNIST 0/1 show these baselines are undertrained. So the reported 56%, 57%, and 36% improvements are not evidence of any quantum advantage. The hardware evaluation uses 8 samples per task without error bars, and the simulated-noise claims are given without error bars either.\n\nWhat is left is a plausible engineering recipe for analog QML. Someone working on Rydberg classifiers may want to borrow the pulse parameterization and the local/global encoding split. But the paper oversells it as a first quantum ResNet. The authors should either supply a genuine derivation of the residual structure or drop that language, and they must redo the classical baseline comparison with proper tuning, multiple seeds, and error bars. The footnote stating that the paper is already published at ICCV 2025 also needs to be resolved if this is being submitted elsewhere.\n\nThis is a paper that deserves referee time—the engineering is concrete and reproducible—but my verdict would be major revision, not acceptance. I'd send it to reviewers with explicit instructions to check the residual mapping and the baseline fairness.","headline":"Useful Rydberg pulse-encoding scheme, but the quantum-ResNet framing and the outperformance claim don't survive close reading.","tokens_in":14800,"tokens_out":4215,"would_cite":false,"duration_ms":40426,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","68T05"],"pacs":["03.67.Ac","03.67.Lx"],"model":"deepseek-v4-flash","headline":"RESQ maps residual networks onto Rydberg atoms and claims a 56% edge over classical peers.","keywords":["quantum machine learning","Rydberg atom quantum computer","residual neural network","neural ordinary differential equations","analog quantum computing","Hamiltonian parameterization","MNIST classification","noise robustness"],"falsifier":"Train the same classical feedforward, residual, and neural-ODE classifiers on the same five PCA features but let them run to convergence (for example, 1,000 or 10,000 Adam iterations instead of 75) and compare final test accuracy; if any reaches or exceeds RESQ's reported accuracy, the claimed 56%, 57%, and 36% improvements are artifacts of the training-time cutoff rather than a property of the quantum model.","tokens_in":13696,"feed_emoji":"⛛️","tokens_out":7098,"duration_ms":69819,"temperature":0.7,"pith_summary":"RESQ introduces a framework for implementing residual neural networks on analog Rydberg-atom quantum computers by encoding input features and trainable parameters directly into the Hamiltonian pulses that drive the atoms. The paper claims this is the first fully quantum-native realization of ResNets, and reports that on binary classification tasks from MNIST, FashionMNIST, and a diabetes dataset it outperforms similarly sized classical feedforward, residual, and neural-ODE networks by 56%, 57%, and 36%, respectively, while remaining robust to hardware noise on QuEra's Aquila device. The motivation is that ResNets have a continuous-time ODE structure that matches the natural Schrödinger dynamics of analog quantum hardware, whereas gate-based quantum computers can only approximate such dynamics with many discrete steps. If the claims hold, RESQ offers a parameter-efficient recipe for building residual architectures natively on near-term analog quantum machines.","feed_headline":"Rydberg-atom framework claims 56% edge in ResNet classification","feed_subtitle":"A 20-parameter analog quantum classifier trained on five PCA features beats feedforward, residual, and neural-ODE baselines.","key_machinery":"The load-bearing object is the parameterized analog Hamiltonian $$H(t)=\\frac{\\$\\Omega$(t)}{2}\\sum_i(e^{i\\$\\varphi$(t)}|g\\rangle_i\\langle r|_i+e^{-i\\$\\varphi$(t)}|r\\rangle_i\\langle g|_i)-\\$\\Delta$(t)\\sum_i \\hat n_i+\\sum_{i<j}\\frac{C_6}{|\\vec p_i-\\vec p_j|^6}\\hat n_i\\hat n_j-\\delta(t)\\sum_i h_i \\hat n_i,$$ with $\\Omega(t)$, $\\Delta(t)$, and $\\delta(t)$ specified as piecewise-linear pulses. RESQ's central parameterization sets each pulse strength at a holding time to $\\theta_j \\omega_i+\\theta_{j+1}$, so a single input feature $\\omega_i$ is scaled and offset by learned parameters throughout the evolution, while alternating local couplings $h_i$ lift further features and parameters into the computation. This turns the hardware's continuous-time evolution into the residual/neural-ODE structure: the input is not fed through fixed layers but shapes the entire state trajectory, and the final averaged $\\langle|1\\rangle$ probability is the classifier output. The stochastic pulse-gradient method supplies unbiased gradients of this analog program with respect to the Hamiltonian parameters.","core_discovery":"RESQ claims that an analog Rydberg-atom quantum computer can implement residual neural networks directly, without gates, by treating the Schrödinger evolution itself as a trainable neural ODE. The framework encodes each input feature $\\omega_i$ into the pulse heights of the global Rabi frequency, global detuning, and local detuning, with each height written as $\\theta_j\\omega_i+\\theta_{j+1}$ for learned parameters $\\theta_j, \\theta_{j+1}$, and encodes additional features and parameters into alternating site-dependent couplings $h_i$ of the local detuning term. After evolving the $N$-atom system for a fixed time, the averaged probability of measuring atoms in the $|1\\rangle$ state is the soft label, and the parameters are trained by backpropagating cross-entropy loss through an unbiased stochastic pulse-gradient estimator. With $N=4$ atoms and three pulse intervals, this gives 20 trainable parameters and five input features; on binary MNIST, FashionMNIST, and diabetes tasks the paper reports 56%, 57%, and 36% accuracy improvements over similarly sized classical feedforward, residual, and neural-ODE networks, respectively, with accuracy and F1 within about 1% of ideal when pulse and position noise is simulated. Real-hardware inference on Aquila is reported to be largely robust except for samples near the decision boundary.","pith_inferences":["In our reading, the reported advantage is more plausibly a statement about baseline starvation than about quantum expressivity: classical architectures trained on the full images, or on the five PCA features with many more iterations, would likely close most of the gap. We would test this before treating the 56% figure as a quantum speedup.","The residual block analogy is conceptual, not literal: Schrödinger evolution is unitary, so RESQ implements a continuous-time integration of a linear operator rather than the additive nonlinear skip layer of the original ResNet; the natural frame for the result is analog quantum neural ODEs.","A direct scaling test would run RESQ with $N>4$ atoms on Aquila to see whether the linear feature-scaling claim survives real interaction graphs and noise; the paper's hardware evaluation is limited to the same $N=4$ configuration used in simulation.","Because the gradient estimator is unbiased and hardware-agnostic, the same piecewise-pulse parameterization could be transferred to other controllable analog platforms, so the framework's value may outlive Rydberg-specific hardware."],"forward_implications":["RESQ gives current analog hardware a usable classification recipe: with four atoms it uses 20 trainable parameters and five input features, and inference on Aquila is reported to stay within about 1% of ideal accuracy under simulated noise.","Because features enter through pulse heights and local couplings, the number of input features grows linearly with qubit count ($3+N/2$ for $N$ atoms) while the evolution time stays fixed, so larger atom arrays could handle richer inputs without deeper circuits.","Residual structure is realized natively as continuous Hamiltonian evolution, which means analog Rydberg systems can bypass the discrete-step decomposition that gate-based quantum ResNet proposals require.","The framework is not tied to images: the same pipeline is demonstrated on a tabular healthcare dataset, suggesting generality across binary classification tasks.","If the results hold, RESQ is the first fully quantum-native neural-ODE classifier, and the mapping from model parameters to Hamiltonian pulses becomes a general template for other continuous-time models."],"supporting_citations":[{"why":"Supplies the hardware model: QuEra Aquila's 256-qubit analog Rydberg computer with global and local control used for real-machine inference.","marker":"[39]"},{"why":"Provides the differentiable analog quantum computing method (stochastic pulse gradient) that RESQ adapts to compute training gradients.","marker":"[23]"},{"why":"Introduces the stochastic parameter-shift rule that underlies the gradient estimation used to train the Hamiltonian parameters.","marker":"[5]"},{"why":"Defines neural ODEs, the continuous-time residual model RESQ claims to realize natively on Rydberg hardware.","marker":"[9]"},{"why":"Supplies the Gaussian-noise model of pulse and position errors used to simulate Aquila's hardware noise.","marker":"[26]"},{"why":"Defines ResNets and residual blocks, the classical architecture RESQ is extending to analog quantum systems.","marker":"[16]"},{"why":"Defines augmented neural ODEs, the classical neural-ODE baseline structure used for comparison.","marker":"[14]"},{"why":"Provides the MNIST dataset used for the pairwise classification benchmarks.","marker":"[11]"},{"why":"Provides FashionMNIST, the second vision benchmark used in evaluation.","marker":"[40]"}],"fun_headline_variants":["ResQ: Rydberg atoms run ResNets via neural ODEs","Analog Rydberg quantum computer trains residual nets directly","Quantum neural ODE on Rydberg atoms beats classical baselines","No gates needed: Rydberg atom evolution does ResNet training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline 'similarly sized classical models' comparison assumes that a classical network trained on the same five PCA features for the same 75 iterations is a meaningful baseline; if classical models are instead given raw data, more features, or longer training, RESQ's reported improvements are not established.","fun_headline_variants_meta":{"raw":{"variants":["ResQ: Rydberg atoms run ResNets via neural ODEs","Analog Rydberg quantum computer trains residual nets directly","Quantum neural ODE on Rydberg atoms beats classical baselines","No gates needed: Rydberg atom evolution does ResNet training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000295,"raw_usage":{"total_tokens":1712,"prompt_tokens":938,"completion_tokens":774,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":699}},"tokens_in":554,"tokens_out":774,"duration_ms":7119,"temperature":1.0,"reasoning_tokens":699,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:23:51.761856+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same classical feedforward, residual, and neural-ODE classifiers on the same five PCA features but let them run to convergence (for example, 1,000 or 10,000 Adam iterations instead of 75) and compare final test accuracy; if any reaches or exceeds RESQ's reported accuracy, the claimed 56%, 57%, and 36% improvements are artifacts of the training-time cutoff rather than a property of the quantum model.","supporting_citations":[{"cited_title":"Differentiable analog quantum computing for opti- mization and control","cited_arxiv_id":null,"evidence_quote":"Provides the differentiable analog quantum computing method (stochastic pulse gradient) that RESQ adapts to compute training gradients."},{"cited_title":"Measuring analytic gradients of general quantum evolution with the stochastic parameter shift rule","cited_arxiv_id":null,"evidence_quote":"Introduces the stochastic parameter-shift rule that underlies the gradient estimation used to train the Hamiltonian parameters."},{"cited_title":"Neural ordinary differential equa- tions","cited_arxiv_id":null,"evidence_quote":"Defines neural ODEs, the continuous-time residual model RESQ claims to realize natively on Rydberg hardware."},{"cited_title":"Digital-analog quantum learning on Rydberg atom arrays","cited_arxiv_id":"2401.02940","evidence_quote":"Supplies the Gaussian-noise model of pulse and position errors used to simulate Aquila's hardware noise."},{"cited_title":"Deep residual learning for image recognition","cited_arxiv_id":null,"evidence_quote":"Defines ResNets and residual blocks, the classical architecture RESQ is extending to analog quantum systems."},{"cited_title":"Aug- mented neural odes","cited_arxiv_id":null,"evidence_quote":"Defines augmented neural ODEs, the classical neural-ODE baseline structure used for comparison."},{"cited_title":"The MNIST Database of Handwritten Digit Images for Machine Learning Research","cited_arxiv_id":null,"evidence_quote":"Provides the MNIST dataset used for the pairwise classification benchmarks."}],"review_version":1}