{"id":"048813ac-1b38-4e80-a72e-f236ad839432","arxiv_id":"2412.01038","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An RL/GNN-based compiler (RLGS) finds emitter-based photonic graph-state generation sequences that reduce generation time by up to 57.5%, emitters by up to 17.5%, and CZ gates by up to 57.8% versus a Stabilizer Solver baseline.","lead":"The paper presents RLGS, a reinforcement-learning and graph-neural-network compiler that chooses sequences of emitter operations for generating photonic graph states, reducing generation time, emitter count, and CZ gate count versus a Stabilizer Solver baseline. It matters because these three costs drive decoherence and gate errors in photonic measurement-based quantum computers, so better generation sequences could improve fidelity.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported metric reductions rely on a 100-random-order Stabilizer Solver baseline, not the exhaustive search the paper invokes; against a stronger baseline the claimed gains may largely disappear.","rationale":"The single most load-bearing assumption behind the abstract's quantitative claim is that the baseline represents a strong or standard comparator. The paper's own description contradicts itself: Section 6.1.2 samples 100 random emission orders, while Section 6.2 credits the baseline with exhaustive search and O(V!·V^4) complexity. The evaluated baseline is thus a weak random heuristic. Because the Stabilizer Solver minimizes N_e only for a fixed emission order and ignores T_gen and N_CZ, the only way RLGS can show large reductions in all three metrics is if the baseline is far from the true optimum. With 100 samples out of V! orders, that is essentially guaranteed for large V. The N_e reduction (13.9-17.5%) is particularly telling: if the baseline truly achieved near-minimal N_e, RLGS should rarely beat it; the fact that it does suggests the baseline's N_e is not near-optimal. The 1200-photon 'reduction=1' artifact confirms the baseline's fragility. This concern is more directly about the central claim than the reader's error-model concern, because the abstract claims metric reductions, not end-to-end fidelity; but it also matters for fidelity, since inflated metric reductions would not translate into fidelity gains. A concrete re-evaluation with a stronger baseline would settle whether the numbers hold.","tokens_in":21056,"tokens_out":4560,"duration_ms":39547,"concrete_test":"For each small/medium benchmark, replace the 100-random-order baseline with (a) an exhaustive enumeration of all emission orders for graphs with V<=20 (or a branch-and-bound search) and (b) for larger graphs, at least 10^5 random orders or a metric-specific search minimizing T_gen, N_e, and N_CZ separately. Recompute the reduction ratios in Fig. 7; if the T_gen reductions drop below roughly 10% or the N_e reductions become negative, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (31.1/49.6/57.5% T_gen reductions, etc.) is measured against a baseline that Section 6.1.2 defines as the Stabilizer Solver run with only 100 random emission orders per graph, taking the best result. This is not the exhaustive-search Stabilizer Solver the paper later invokes in Section 6.2, which it credits with O(V!·V^4) complexity. For hwea-6 (V=26), 100 samples out of 26! orders is negligible; for larger graphs the coverage is vastly smaller. Since the Stabilizer Solver only minimizes N_e for a given emission order and does not optimize T_gen or N_CZ, the baseline is expected to be far from optimal on those metrics, and 100 random orders also makes its N_e nearly arbitrary. The reported reductions may therefore reflect baseline weakness rather than RLGS superiority. The scalability section compounds this: when the baseline fails to find any sequence at 1200 photons, reductions are set to 1 (100%), which inflates the comparison. A fair test requires an exhaustive or much stronger baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RLGS, a reinforcement-learning framework combined with graph neural networks, to construct emitter-based photonic graph-state generation sequences. RLGS selects among six graph operations in a backward search, with a reward function that penalizes generation time and, through an emitter-swap penalty, also targets the number of emitters and CZ gates. The authors train a single Q-network on small benchmarks and evaluate it on 18 graph states from six quantum applications, reporting average reductions in generation time of 31.1%, 49.6%, and 57.5% for small, medium, and large graphs, with corresponding reductions in emitter count and CZ gates, compared to a Stabilizer Solver baseline.","tokens_in":21270,"tokens_out":5306,"duration_ms":50390,"significance":"If the quantitative claims were supported, RLGS would be a valuable compilation technique: the graph-operation formulation avoids the O(V!·V^4) exhaustive search, the receptive-field mechanism gives a practical O((V+E)·W) inference procedure, and the trained Q-network is intended to generalize across graph sizes and applications. The paper also makes a useful contribution by explicitly modeling three fidelity-relevant metrics rather than emitter count alone. However, the current evaluation does not establish the headline reductions: the baseline is only 100 random emission orders, three training graphs appear in the test set, the 1200-photon result converts a baseline failure into a 100% reduction, and the reward is never validated against the paper's own fidelity model. These issues affect the central claim and prevent the results from being accepted as reported.","major_comments":[{"comment":"The baseline is not the exhaustive Stabilizer Solver described in Section 4.1. It is the Stabilizer Solver run on 100 random emission orders with the best result selected. For hwea-6 (V=26), 100 samples out of 26! orders is negligible, and for larger graphs such as qft-14 (V=235) the coverage is even smaller. Because the Stabilizer Solver minimizes only N_e for a fixed emission order and does not optimize T_gen or N_CZ, the T_gen and N_CZ values from the baseline are essentially arbitrary for the sampled orders. The reported reductions of 31.1%/49.6%/57.5% in T_gen and 37.7%/53.4%/57.8% in N_CZ are therefore likely inflated by baseline weakness. I ask the authors to compare against exhaustive search on small graphs where feasible and to report absolute metric values, not only reduction ratios.","section":"Section 6.1.2"},{"comment":"The Q-network is trained on hwea-6, hc-6, and qft-5, and then evaluated on all 18 benchmarks in Table 1, including those same three training graphs. The small-size averages in Fig. 7 are therefore partially in-sample. Section 6.5 repeats this issue: each of the 20 training combinations evaluates on the three graphs used in that combination. The claim of generalization to unseen graph states requires a held-out evaluation that excludes training graphs, or at least an explicit separation of in-sample and out-of-sample results.","section":"Section 6.1.4 / Section 6.2"},{"comment":"Setting the reduction ratios to 1 for 1200-photon graphs because the baseline \"fails to find a solution\" is not a valid metric comparison. A baseline that cannot produce a valid sequence within its sampling budget should be reported as a failure rate, not as a 100% reduction. This is especially problematic because the baseline uses only 100 random emission orders; its failure at large sizes is expected and says little about the quality of RLGS sequences. Please report baseline success rates and, if possible, use a baseline that completes on these graph sizes.","section":"Section 6.4"},{"comment":"The reward function applies an Emitter-Swap penalty based on the claim in Lemma 5.1 that N_CZ is proportional to N_e. The lemma's derivation assumes that the average number of edges removed per absorption (β) and per CZ gate (γ) are constants, but these averages depend on the actual operation sequence, so the proportionality is not established. More importantly, the reward is never connected to the fidelity model of Section 3: no experiment reports F = exp(-N_e·T_gen/T_2)·σ_CZ^{N_CZ} or the photon-loss probability. Since the paper motivates all three metrics through execution fidelity, the reader cannot determine whether the reported metric reductions improve fidelity. Please add a fidelity comparison using the Section 3 model, or clearly state that fidelity is not claimed.","section":"Section 5.1.3 / Lemma 5.1"},{"comment":"The description of the baseline says it reports the generation sequence that yields the \"best result (in terms of the three metrics),\" but no selection criterion is specified. If the baseline is selected, for example, by minimal T_gen, then its N_e and N_CZ values need not be representative of the baseline method. Please specify the exact selection rule (e.g., lexicographic minimization of the three metrics in a fixed order) and report the distribution of baseline metrics over the 100 random orders.","section":"Section 6.1.2"}],"minor_comments":[{"comment":"The phrase \"Reinforcement Larning-guided Graph State generation\" should be corrected to \"Reinforcement Learning-guided Graph State generation\".","section":"Abstract / Section 1"},{"comment":"The y-axis label \"Relative Values and Time Improvement\" is ambiguous; please define whether higher values are better and specify the units for the time improvement.","section":"Section 6.3.2, Fig. 9"},{"comment":"The five graph state sizes of 400, 600, 800, 1000, and 1200 photons are not associated with any graph family or generation procedure; without this information the scalability results cannot be reproduced.","section":"Section 6.4"},{"comment":"The paper lists gate times and coherence parameters but does not state whether T_gen includes only quantum gate times or also emission and measurement overheads; please clarify the exact calculation of T_gen.","section":"Section 6.1.3"},{"comment":"The column headers \"V E Graph state size\" should be separated into clearly labeled columns, and the table should have a caption.","section":"Table 1"},{"comment":"References [22] and [52] are the same paper by Bremner, Jozsa, and Shepherd and should be merged.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a timely and interesting problem, and the RL/GNN approach is plausible, but the headline numbers are not yet supported by the evaluation. The weak 100-random-order baseline, the inclusion of training graphs in the test set, and the conversion of baseline failure into 100% reductions are all fixable with additional experiments. I would encourage the editor to send the paper back for major revision rather than reject it, because the central ideas could be validated with a stronger baseline, a proper held-out split, and fidelity reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper puts a DQN/GNN agent on the six emitter operations from Kaur et al., with a reward that jointly penalizes generation time, emitter count, and CZ gates. That is a genuine gap: prior work like the Stabilizer Solver only minimizes emitter count. The framework is clearly described, the receptive-field trick is a practical way to cut the action space, and the complexity analysis is honest. If the method works, it is a useful compiler step for emitter-based MBQC and repeaters.\n\nBut the headline numbers should not be taken at face value. The baseline is the Stabilizer Solver with only 100 random emission orders per graph, taking the best. The paper later invokes the exhaustive solver with O(V!·V^4) complexity to argue the search space is large. For a 26-photon graph, 100 samples out of 26! is minuscule, so the reported 31.1–57.5% reductions in T_gen are mostly an artifact of the weak baseline. The 1200-photon result sets reductions to 100% because the baseline fails to find any sequence; that is not a fair comparison, and it inflates the scalability story.\n\nThere is also leakage: the main results train on hwea-6, hc-6, and qft-5, then report small-graph results on those same three graphs. The held-out comparison should exclude training graphs or report them separately. No error bars appear anywhere, and the reward is a proxy combination that is never checked against the actual fidelity model from Section 3, so the link from metric reductions to fidelity is asserted rather than demonstrated.\n\nThat said, the idea is worth building on. The paper deserves a serious referee, but the authors should be asked to compare against a much stronger baseline (exhaustive or a large random/greedy search), to report variance across seeds and Q-network initializations, and to present true held-out results. With those changes, the method may well stand up; without them, the current claims are not credible.\n\nI would send it to review rather than desk-reject, with the baseline issue as the primary revision request.\n\nBest,","headline":"A promising RL/GNN approach to emitter-based photonic graph-state generation, but the headline reductions rest on a 100-random-order baseline and some train/test leakage.","tokens_in":21790,"tokens_out":2757,"would_cite":false,"duration_ms":24671,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.Lx","42.50.Ex"],"model":"deepseek-v4-flash","headline":"A single reinforcement-learning agent can choose emitter sequences that cut photonic graph-state generation time by up to 57.5 percent.","keywords":["photonic quantum computing","graph states","quantum emitters","reinforcement learning","graph neural networks","measurement-based quantum computation","compilation"],"falsifier":"Simulate or run the RLGS-selected and Stabilizer-Solver-selected sequences on an emitter platform with a full error model including decoherence, CZ gate errors, and photon loss in delay lines, then measure the actual fidelity or entanglement visibility of the generated graph states; if sequences with larger $T_{gen}$, $N_e$, or $N_{CZ}$ achieve equal or better fidelity, the paper's central claim that minimizing these proxies improves fidelity is falsified. A cheaper check is to sweep alpha and confirm that measured fidelity tracks the reward direction predicted by the paper's error model.","tokens_in":20864,"feed_emoji":"⚛️","tokens_out":5662,"duration_ms":48022,"temperature":0.7,"pith_summary":"This paper argues that the order of operations used to build a photonic graph state from quantum emitters can be chosen by a reinforcement-learning agent that jointly optimizes all three metrics known to affect fidelity: generation time, number of emitters, and number of CZ gates. Earlier work, the Stabilizer Solver, only minimized emitter count and had to search over $V!$ emission orders. RLGS trains a GNN-based Q-network once on small graph states, then applies it to unseen graphs of up to 1200 photons. Compared with the baseline, the paper reports average generation-time reductions of 31.1%, 49.6%, and 57.5% for small, medium, and large graph states, with emitter reductions of 13.9%, 16.7%, and 17.5% and CZ-gate reductions of 37.7%, 53.4%, and 57.8%. If these proxy-metric gains translate through the paper's error model, RLGS would directly improve the fidelity of measurement-based photonic computation.","feed_headline":"RL agent cuts photonic graph-state generation time up to 57.5%","feed_subtitle":"One trained Q-network beats emitter-only optimization across unseen graph sizes, trimming CZ gates too.","key_machinery":"The load-bearing machinery is the backward graph-operation model: six deterministic operations (Emitter Swap, Type-I/II/III absorption, Reversed CZ, and Type-III reversed CZ) that each remove a photon or edge and map to concrete quantum gates, so any action sequence is a valid generation sequence. On top of it, RLGS trains a deep Q-network with two GIN layers followed by three MLP layers, experience replay, a target network, and epsilon-greedy exploration. The reward is $-\\mathrm{add}(T_{gen})$ for ordinary actions and $-\\mathrm{add}(T_{gen}) - \\alpha T_{CZ}$ for Emitter Swap, with $\\alpha$ a user-set trade-off parameter. The receptive field (default $W = 0.5V$) restricts each decision to the closest $W$ photons, cutting inference complexity to $O((V+E)W)$ versus $O(V! \\cdot V^4)$ for the baseline.","core_discovery":"The central claim is that a generation sequence for an emitter-based photonic graph state can be treated as a Markov decision process over six backward graph operations, and that a deep Q-network with a graph-neural-network encoder can learn a policy whose choices dominate those of the only prior solver. RLGS operates backward: starting from the target graph, it removes photons and edges using Emitter Swap, three absorption types, and reversed CZ operations, then reverses the sequence into forward gates. The reward penalizes added generation time, and penalizes Emitter Swap with a fraction of CZ-gate time, under the lemma that CZ count is proportional to emitter count for fixed photon and edge counts. At inference, a receptive field of half the photon count limits the action space. The paper's reported result is that this single trained network reduces $T_{gen}$ by 31.1% to 57.5%, $N_e$ by 13.9% to 17.5%, and $N_{CZ}$ by 37.7% to 57.8% on average across six benchmark applications, with reductions growing with graph size and with the baseline failing at 1200 photons while RLGS still succeeds.","pith_inferences":["Editorial inference: the reported gains are on proxy metrics, not measured fidelity; translating them into guaranteed fidelity improvement requires validating the paper's decoherence and CZ-error model on the target hardware, since a misweighted alpha could in principle pick a sequence that scores better on all three proxies yet is no better in actual error.","Editorial inference: because the action set is hardware-agnostic and the reward weights are parameters, the same trained framework could be adapted to other emitter platforms by retraining or reweighting alpha, which would test whether the learned policy transfers across technologies.","Editorial inference: the baseline comparison uses the best of 100 random emission orders rather than exhaustive search; on small graphs, an exhaustive Stabilizer Solver comparison would establish how much of the reported gap is due to RLGS versus a weak baseline.","Editorial inference: RLGS optimizes only the generation stage and explicitly leaves measurement-stage photon loss aside; combining it with measurement-loss-aware compilation could compound fidelity gains, but the combined effect is an open question."],"forward_implications":["A single offline-trained Q-network replaces per-graph exhaustive search: RLGS infers generation sequences for unseen benchmarks without retraining, and the paper reports robust results across all 20 three-benchmark training combinations.","The method scales to graph states where the baseline breaks down: the Stabilizer Solver finds no valid sequence at 1200 photons, while RLGS returns one for every tested size.","Reduction ratios grow with graph size, from 31.1% to 57.5% for generation time, so the advantage compounds as the search space grows factorially.","The alpha knob gives users a fidelity trade-off: alpha = 0.1 favors speed with generation-time reductions up to 75.5% at the cost of more emitters, while alpha = 1.0 favors fewer emitters and CZ gates.","A half-size receptive field preserves nearly all quality while making inference 1.45 times faster, and a 0.05V field gives a 13 times speedup at a measured quality cost."],"supporting_citations":[{"why":"Defines the emitter-number-minimizing Stabilizer Solver and the $O(V! \\cdot V^4)$ search complexity that RLGS uses as its baseline.","marker":"[9]"},{"why":"Supplies the six backward graph operations that form RLGS's action space.","marker":"[10]"},{"why":"Provides the decoherence fidelity model $F_{de} = e^{-N_e T_{gen}/T_2}$ used to justify optimizing generation time and emitter count.","marker":"[8]"},{"why":"Establishes the CZ gate error model and resource requirements for all-photonic graph states, grounding the $N_{CZ}$ metric.","marker":"[33]"},{"why":"Gives quantum-dot emitter parameters and the single-CZ fidelity $\\sigma_{CZ} = 99\\%$ used in evaluation.","marker":"[34]"},{"why":"Introduces the deep Q-network training procedure with experience replay and target network that RLGS adopts.","marker":"[6]"},{"why":"Provides GIN, the graph neural network architecture used to embed graph states for the Q-network.","marker":"[23]"}],"fun_headline_variants":["RL cuts photonic graph-state generation time by up to 57%","RL agent reduces photonic graph-state time, emitters, and CZ gates","RL-trained compiler scales photonic graph generation beyond 1200 photons","Deep Q-network beats old solver on photonic graph-state metrics","Backward RL sequence trims time, emitters, and CZ gates in photonic states"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the three proxy metrics—generation time, emitter count, and CZ-gate count—capture the errors that dominate real graph-state fidelity, so the reward function's weighted combination tracks fidelity; if other errors dominate or the user-set alpha mismatches the hardware, smaller metrics need not mean a better graph state.","fun_headline_variants_meta":{"raw":{"variants":["RL cuts photonic graph-state generation time by up to 57%","RL agent reduces photonic graph-state time, emitters, and CZ gates","RL-trained compiler scales photonic graph generation beyond 1200 photons","Deep Q-network beats old solver on photonic graph-state metrics","Backward RL sequence trims time, emitters, and CZ gates in photonic states"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000669,"raw_usage":{"total_tokens":3081,"prompt_tokens":1004,"completion_tokens":2077,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":1977}},"tokens_in":620,"tokens_out":2077,"duration_ms":13111,"temperature":1.0,"reasoning_tokens":1977,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:44:38.496311+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate or run the RLGS-selected and Stabilizer-Solver-selected sequences on an emitter platform with a full error model including decoherence, CZ gate errors, and photon loss in delay lines, then measure the actual fidelity or entanglement visibility of the generated graph states; if sequences with larger $T_{gen}$, $N_e$, or $N_{CZ}$ achieve equal or better fidelity, the paper's central claim that minimizing these proxies improves fidelity is falsified. A cheaper check is to sweep alpha and confirm that measured fidelity tracks the reward direction predicted by the paper's error model.","supporting_citations":[{"cited_title":"Economou, and Edwin Barnes","cited_arxiv_id":null,"evidence_quote":"Defines the emitter-number-minimizing Stabilizer Solver and the $O(V! \\cdot V^4)$ search complexity that RLGS uses as its baseline."},{"cited_title":"Economou, and Shuo Sun","cited_arxiv_id":null,"evidence_quote":"Provides the decoherence fidelity model $F_{de} = e^{-N_e T_{gen}/T_2}$ used to justify optimizing generation time and emitter count."},{"cited_title":"Economou","cited_arxiv_id":null,"evidence_quote":"Establishes the CZ gate error model and resource requirements for all-photonic graph states, grounding the $N_{CZ}$ metric."},{"cited_title":"Economou","cited_arxiv_id":null,"evidence_quote":"Gives quantum-dot emitter parameters and the single-CZ fidelity $\\sigma_{CZ} = 99\\%$ used in evaluation."}],"review_version":1}