{"id":"2d976677-6e2f-4ec7-af6d-7412301b54a7","arxiv_id":"2607.26438","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A spectrum-resolved HOM interference readout, mapped into an actor-critic, is claimed to outperform matching MLP agents on continuous-control benchmarks and to restore drifted transmon-gate fidelities in simulation.","lead":"This paper proposes a photonic reinforcement-learning design that reads out frequency-resolved Hong-Ou-Mandel interference as a tensor instead of a single coincidence count, using the tensor's diagonal entries as continuous actions and its correlations as value features. In simulated benchmarks, it reports faster and stabler learning than parameter-matched neural networks, and simulates the same architecture correcting drift in superconducting CZ and iSWAP gates toward their","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The finite-bin discretization of Eq. (2) makes all diagonal C_ii zero, so the actor readout in Eqs. (4)-(5) would be constant; the SM never specifies the actual C_ij formula, leaving the central benchmarks unverifiable.","rationale":"The reader's weakest assumption—that the simulations use an exact-probability oracle with no photon-counting noise—is valid but secondary. The more load-bearing issue is that the simulated C_ij is not actually specified, and the natural finite-bin discretization of the paper's own Eq. (2) has zero diagonal entries, which would make the actor output constant. Since the paper's empirical claims all depend on training this actor, the absence of a defined, reproducible C_ij formula means the central results cannot be checked. This should be the primary condition for acceptance: the authors must disclose and justify the exact numerical tensor and control for the action-space constraint. If the disclosed implementation turns out not to match Eq. (2), the central claim collapses; if it does match via some sub-bin construction, the finite-shot oracle concern raised by the reader remains an additional but less fundamental issue. A conditional verdict is therefore appropriate: the paper is not acceptable as-is, but it may become acceptable if the numerical definition is supplied and the benchmark comparison is rerun with an action-space-matched MLP control.","tokens_in":19152,"tokens_out":15787,"duration_ms":164094,"concrete_test":"Obtain the authors' code (or the exact finite-bin expression for C_ij) and reimplement the numerical SR-HOM module as specified in the SM. With random normalized spectral vectors ψ and λ, compute the diagonal entries C_ii using the discrete form of Eq. (2). If all C_ii are zero, then the reported actor is constant and the learning curves cannot originate from Eqs. (2)-(5). If C_ii are nonzero, inspect the formula: check whether it is a different, unpublished model and, as a control, run the MLP baseline with its actions projected onto the polytope Σ_r g_r^{-1}(a_r) ≤ 1 implied by the normalization; if the constrained MLP matches SR-HOM, the action-space restriction rather than the optical features explains the gains.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central numerical object is the spectrum-resolved coincidence tensor C_ij, but the Supplemental Material does not give the finite-bin formula used to compute it. It says only that C_ij is 'evaluated directly from the normalized spectral-mode coefficients in the exact-probability limit.' If the implementation is the natural discrete version of Eq. (2), with one complex amplitude per frequency bin, then for i=j the two-photon amplitude is ψ_i φ_i − ψ_i φ_i = 0, so every diagonal entry C_ii is exactly zero. Equation (4) then gives p_i=0 and Eq. (5) makes every action a_r = g_r(0) constant, which contradicts the reported learning curves on all five benchmarks. Thus either the numerical C_ij is not the discrete version of Eq. (2), or the SM omits a sub-bin integration or another modification that makes the diagonal nonzero. If the implementation instead uses a product form |ψ_i|^2|λ_j|^2 or another ad hoc spectral correlation, the benchmark results do not test the claimed SR-HOM/HOM architecture. This issue is prior to the finite-shot oracle concern: even with exact probabilities, the object being optimized may not be the physical HOM tensor. Additionally, if C_ii were nonzero, Eq. (3)'s normalization Σ_{ij}C_ij=1 implies Σ_r z_r = trace(C) ≤ 1, so the mean action is confined to a simplex-constrained subset of the action box; for Ant-v5 (m=8, actions in [-1,1]) the feasible region has volume fraction far below 10^{-4}, an unacknowledged restriction relative to the full-box MLP baseline.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a spectrum-resolved Hong-Ou-Mandel (SR-HOM) architecture for reinforcement learning, in which the frequency-resolved coincidence tensor C_ij of two interfering single photons is used as a feature map: grouped diagonal entries produce continuous actions, while the tensor (or a related readout) supplies state-action value features. The authors report that a parameter-matched MLP actor-critic is outperformed on five Gymnasium continuous-control benchmarks, with a 4.4x improvement in episodes-to-threshold on LunarLanderContinuous-v3, and that the same architecture learns an online calibration policy that partly recovers drifted CZ and iSWAP gate fidelities. The central numerical experiments use an exact-probability limit for C_ij, as explicitly stated in the Supplemental Material.","tokens_in":19609,"tokens_out":7891,"duration_ms":89046,"significance":"The underlying idea — using the spectral structure of HOM interference as a trainable tensor feature map rather than a scalar overlap — is original and could be of interest to the quantum-optical machine-learning community. The paper is also honest in disclosing that the reported learning curves exclude finite-shot photon-counting fluctuations, and the tunable-coupler calibration environment is described in considerable detail. However, the central numerical object of the paper, the finite-bin C_ij actually computed in the simulations, is never specified. As written, the paper does not provide enough information to verify the benchmark results, and at least one natural implementation of the stated equations makes the actor readout identically constant. Until the numerical model is fully specified and the reported results are reproduced with that model, the empirical claims cannot be interpreted as evidence for the SR-HOM architecture.","major_comments":[{"comment":"The numerical formula for C_ij is never given. The SM says only that C_ij is 'evaluated directly from the normalized spectral-mode coefficients in the exact-probability limit,' with one complex amplitude per frequency bin. If the spectral amplitude is piecewise constant within each bin, then for i=j the integrand in Eq. (2) is ψ_i φ_i − ψ_i φ_i = 0, so every diagonal entry C_ii is exactly zero. Eq. (5) then makes all actions constant, contradicting every reported learning curve. If a sub-bin integration, a different basis, or a modified tensor was used, that formula must be stated explicitly. This is a load-bearing omission: the central benchmarks depend on the value of C_ii, and the paper currently leaves that object undefined.","section":"SM 'Numerical Implementation', Eq. (S2); main Eqs. (3)-(5)"},{"comment":"The reported 4.4x 'sample efficiency' is episodes-to-threshold in an exact-probability oracle, not a hardware sample-efficiency result. The SM explicitly says finite-shot photon-counting fluctuations are not included in the learning curves. Eq. (8) is a Hoeffding bound on estimating the diagonal entries p_i from M accepted coincidence samples; it is not an RL sample-complexity bound and does not imply that 666 training episodes correspond to a comparable physical resource cost. A fair comparison would need to count photon-pair samples per action or provide a rigorous bound converting estimation error into policy performance. As it stands, the headline efficiency claim concerns only a noiseless oracle.","section":"Main 'Complexity Analysis', Eqs. (8)-(9); SM 'Numerical Implementation'"},{"comment":"The critic actually implemented in the SM uses F^Q_i ∝ |α_i|^2 |u_i|^2 |⟨v,β⟩|^2, a separable product of spectral intensities and an overlap. This is not the spectrum-resolved HOM tensor C_ii defined by Eq. (3); it contains no two-photon interference and no coincidence-tensor structure. Consequently, the benchmark results do not evaluate the 'same SR-HOM architecture' claimed for both actor and critic. The paper should either implement the critic with the actual tensor readout or explicitly acknowledge that the critic is a classical surrogate feature map.","section":"SM 'Critic readout', Eq. (S8); main Eqs. (6)-(7)"},{"comment":"Because Σ_{i,j} C_ij = 1 and all entries are nonnegative, Σ_r z_r = trace(C) ≤ 1. Thus the deterministic mean actions are confined to a simplex-constrained subset of the action box. For Ant-v5 (m=8, action bounds [−1,1]), even with an affine map g sending [0,1] to [−1,1], the feasible mean actions satisfy Σ_r (a_r+1)/2 ≤ 1, so most of the action space is inaccessible to the deterministic policy. The paper does not discuss this restriction or its effect on the comparison with an unconstrained MLP. This is especially relevant because the SR-HOM advantage is claimed on exactly these benchmarks.","section":"Main Eqs. (3)-(5); Table S2 (Ant-v5)"}],"minor_comments":[{"comment":"The text contains a recurring typo 'iSW AP' where 'iSWAP' is intended (e.g., in the abstract and Section 'Online RL calibration'). Please correct.","section":"Throughout"},{"comment":"The paper does not report the number of SR-HOM seeds or the distribution of SR-HOM results, while it does show multi-seed MLP curves. Reporting only the best MLP run and a single (or unstated) SR-HOM trajectory makes the comparison difficult to assess. Error bars or multiple SR-HOM seeds should be provided.","section":"Benchmark protocol, SM"},{"comment":"No code or complete hyperparameter list (K, group sizes, learning rates, network widths, seed counts) is provided. Given the missing C_ij formula, a code release or a fully specified implementation appendix is essential for reproducibility.","section":"SM 'Numerical Implementation'"},{"comment":"The gate-calibration results are reported as absolute improvements over the uncorrected baseline, but no MLP or classical controller baseline is included. This makes it difficult to attribute the calibration success specifically to the SR-HOM architecture.","section":"Online RL calibration"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising idea and a detailed physical model, but the central numerical object is unspecified and the implemented critic does not match the claimed architecture. I would not recommend acceptance before the authors provide the exact finite-bin C_ij formula (or code) and address the diagonal-zero issue. The exact-probability-oracle caveat should also be reconciled with the 'sample efficiency' language. These are fixable in a revision, but the current version does not support the reported quantitative claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read. The core idea—use the full spectrum-resolved HOM coincidence tensor as a trainable feature map, with diagonal entries generating continuous actions and off-diagonal terms feeding a value head—is a legitimate step beyond the scalar-overlap HOM neuron. The tunable-coupler calibration application, where the agent sees only measurement diagnostics and fidelity is reserved for validation, is a sensible way to make the proposal concrete. Give credit where it's due: parameter-matching the MLP and picking its best seed is a fair-dealing baseline choice, and the sampling bound in Eq. (8) is correct as far as it goes.\n\nThe problem is that the numerical object at the center of the benchmarks is never pinned down. The SM says C_ij is 'evaluated directly from the normalized spectral-mode coefficients in the exact-probability limit,' but it never gives the finite-bin formula. If you discretize Eq. (2) with piecewise-constant spectral amplitudes—the natural reading—then for i=j the two terms in the amplitude are identical and C_ii=0 exactly. Eqs. (4)–(5) then make every action component constant, which contradicts every reported learning curve. So either the implementation is not the discrete version of Eq. (2), or the SM omits some sub-bin integration or alternative construction. This is not a cosmetic gap: the reader cannot tell whether the benchmarks test a physical HOM tensor at all. The stress-test note has this right, and the paper doesn't answer it.\n\nThe second soft spot is the exact-probability oracle. The learning curves—and the headline 4.4x sample-efficiency number—come from exact coincidence probabilities, not from sampled photon counts. The paper explicitly says finite-shot fluctuations are excluded. The Hoeffding bound in Eq. (8) is the only bridge to hardware, but it is never connected to the episodes-to-threshold numbers. On top of that, there are no error bars or seed-to-seed curves for SR-HOM itself; only the MLP's multi-seed runs are shown.\n\nIs this a desk reject? No. The intended audience—quantum-optical-computing and RL-for-control people—should see this, and a serious referee could help the authors fix the gaps. But the authors need to (1) write down the exact C_ij formula used, (2) inject photon-counting noise into the reported curves, and (3) report seed variance for SR-HOM. Without that, the central comparison is unverifiable. Send it to review, but make those conditions explicit.","headline":"The SR-HOM actor-critic idea is worth a serious look, but the paper never specifies the numerical coincidence tensor it actually optimizes, and the natural discretization would make the action readout constant—so the benchmark story is unverifiable as written.","tokens_in":20135,"tokens_out":3645,"would_cite":false,"duration_ms":45661,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["42.50.-p","03.67.-a"],"model":"deepseek-v4-flash","headline":"Resolving the HOM coincidence spectrum into a tensor readout yields, the paper claims, an optical actor-critic that beats parameter-matched neural baselines and restores drifted two-qubit gates to above 99.8% of reference fidelity.","keywords":["spectrum-resolved Hong-Ou-Mandel interference","quantum optical neural network","reinforcement learning","actor-critic","continuous control","photon frequency encoding","tunable-coupler gate calibration","two-photon interference"],"falsifier":"Re-run the five benchmarks with the exact tensor replaced by empirical estimates from a finite number M of coincidence samples (e.g., M chosen from the paper's bound with ε=0.05, δ=0.01). If the SR-HOM learning curves collapse to or below the parameter-matched MLP baseline once shot noise is present, the claimed advantage is falsified. A hardware experiment that measures the binned coincidence tensor via frequency-resolved detection on a photonic chip would directly test whether the exact-probability oracle is physically achievable.","tokens_in":19044,"feed_emoji":"⚛️","tokens_out":11288,"duration_ms":91827,"temperature":0.7,"pith_summary":"The paper sets out to establish that the spectral structure discarded by standard Hong-Ou-Mandel (HOM) coincidence counting can be promoted into a trainable computational resource. It introduces a spectrum-resolved HOM (SR-HOM) readout that outputs a frequency-binned coincidence tensor, uses grouped diagonal entries to generate continuous actions and the full tensor to estimate action values, and reports that this optical actor-critic learns faster and more stably than parameter-matched multilayer-perceptron baselines on five continuous-control benchmarks. The same architecture is then applied to online calibration of drifted tunable-coupler two-qubit gates, restoring CZ and iSWAP fidelities to 0.9917 and 0.9952, over 99.8% of their drift-free references. The broader claim is that photonic interference can act as a nonlinear feature map embedded in a learning architecture, not merely as a similarity measurement. A sympathetic reader would care because it points toward compact optical hardware that outputs structured continuous signals directly from photon statistics.","feed_headline":"Spectrum-resolved HOM readout tops neural nets on five control tasks","feed_subtitle":"The same frequency-binned photon tensor drives actor and critic, and restores drifted gates to 99.8% of reference fidelity.","key_machinery":"The central object is the spectrum-resolved coincidence tensor C_ij(x,λ), the frequency-binned joint probability that two interfering photons are detected in bins i and j, normalized to sum to one. It replaces the scalar coincidence probability of standard HOM detection and serves as a structured optical feature map: grouped diagonal sums define the action through a fixed monotone rescaling, and the tensor's spectral correlations are passed through a trainable linear readout to produce Q-values. The paper also proves a sampling bound — O(ε^-2 log k) coincidence samples suffice to estimate all diagonal entries to accuracy ε with high confidence, independent of the state dimension — which is t","core_discovery":"The central claim is that replacing the scalar HOM visibility with a spectrum-resolved coincidence tensor turns the interferometer into a trainable feature map for reinforcement learning. Concretely, an input photon whose spectral mode encodes the environment state interferes with a trainable probe photon; the frequency-resolved coincidence density is binned into a k×k tensor C_ij(x,λ). Grouped diagonal sums, passed through a fixed monotone transformation, produce the continuous action vector, while the tensor's spectral correlations are linearly read out to estimate Q-values. The paper reports that on five continuous-control benchmarks this SR-HOM agent reaches reward thresholds faster than","pith_inferences":["If finite-shot photon-counting noise is introduced at the level allowed by Eq. (8), the performance gap versus the MLP baseline may narrow; a natural next step is to re-run the same benchmarks with empirical coincidence samples and measure the degradation.","The diagonal-for-action/off-diagonal-for-value assignment is a general design principle that could transfer to other multi-photon interference platforms, where tensor-shaped readouts might provide different inductive biases than scalar overlaps.","The calibration result hints at a broad recipe: any experimental device with a slow drift and a cheap diagnostic signal could be regulated by an RL agent reading a physical interference tensor, without a model of the underlying drift.","The authors leave open whether the learned calibration policy generalizes to drift excursions larger than those sampled during training; a testable extension is to evaluate on out-of-distribution drift magnitudes."],"forward_implications":["If SR-HOM is correct, a compact photonic device (state-encoded single photon, reconfigurable probe, grating demultiplexing) could implement an actor-critic agent whose action generation cost grows only logarithmically in spectral resolution and not with state dimension.","The same tensor readout can be repurposed beyond RL: the paper argues that assigning distinct computational roles to diagonal, grouped, and off-diagonal components extends naturally to classification, prediction, and other decision-making tasks.","For superconducting processors, the online calibration policy could extend intervals between full recalibrations by tracking slow control-line distortions with low-overhead diagnostics, since the true fidelity is excluded from observations and reward.","The reported stability in a strongly capacity-limited regime suggests physical interference primitives may act as well-behaved function approximators where equally sized classical networks fail to converge."],"fun_headline_variants":["Photon spectrum as trainable feature for quantum RL agents","SR-HOM readout lifts RL sample efficiency over MLP baselines","Quantum optical RL beats neural nets via spectrum-resolved HOM","Frequency-binned HOM tensor enables continuous-action quantum RL","Spectrum-resolved HOM readout tops MLPs on control tasks"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"All reported learning curves and the claimed advantage over the neural baseline assume the spectrum-resolved coincidence tensor is read out as an exact, noiseless probability; the supplemental material explicitly states that finite-shot photon-counting fluctuations are not included in the reported learning curves.","fun_headline_variants_meta":{"raw":{"variants":["Photon spectrum as trainable feature for quantum RL agents","SR-HOM readout lifts RL sample efficiency over MLP baselines","Quantum optical RL beats neural nets via spectrum-resolved HOM","Frequency-binned HOM tensor enables continuous-action quantum RL","Spectrum-resolved HOM readout tops MLPs on control tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000827,"raw_usage":{"total_tokens":3445,"prompt_tokens":729,"completion_tokens":2716,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":2638}},"tokens_in":473,"tokens_out":2716,"duration_ms":22270,"temperature":1.0,"reasoning_tokens":2638,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T15:44:14.423710+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the five benchmarks with the exact tensor replaced by empirical estimates from a finite number M of coincidence samples (e.g., M chosen from the paper's bound with ε=0.05, δ=0.01). If the SR-HOM learning curves collapse to or below the parameter-matched MLP baseline once shot noise is present, the claimed advantage is falsified. A hardware experiment that measures the binned coincidence tensor via frequency-resolved detection on a photonic chip would directly test whether the exact-probability oracle is physically achievable.","supporting_citations":[],"review_version":1}