{"id":"2d3b4641-bd9b-45ca-8eac-f98074c83d95","arxiv_id":"2505.12461","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A DQN-based scheduler for quantum repeater networks trades off request delay against fairness and beats greedy, proportional fair, and FIFO baselines in the authors' simulations.","lead":"This paper trains a Deep Q-Network to schedule which entanglement requests a quantum repeater network serves in each time slot, balancing delay and fairness. The authors report that the learned scheduler matches greedy scheduling on delay while improving fairness, and can be tuned to prioritize fairness.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fairness weight in the DQN reward is action-independent: Eq. 9 makes r2 a constant, so c_j cannot change the scheduler's choices; the claimed delay–fairness trade-off is not implemented as written.","rationale":"The reader's weakest assumption focuses on transferability of the homogeneous-link model, but the more immediate problem is internal: the reward equations in the paper's own method section cannot implement the advertised fairness bias. This is not a question of missing seeds or code; it is a mathematical property of the printed equations. The reader did note that 'the reward uses the same metrics as the evaluation,' which is related, but did not identify that r2 is action-independent. That distinction matters because the central contribution is a tunable delay/fairness scheduler; if the fairness knob is inert, the headline comparison does not demonstrate what is claimed. I therefore recommend REJECT for the paper as written, while noting that a corrected reward that actually includes a state-dependent fairness term could make the claim testable and might restore a conditional acceptance.","tokens_in":7363,"tokens_out":4903,"duration_ms":57573,"concrete_test":"Analytical check: for any state s and two actions a,a', Eqs. 8–10 give ΔQ(s,a,a') = c_d·(r1(s,a) − r1(s,a')) plus terms independent of c_j; hence c_j cancels from every policy decision. Empirically, rerun the Table IV configuration with c_j = 0 and c_j = 0.85 while keeping c_d = 0.15; if Eq. 9 is used, the learned policies and Jain indices should be statistically indistinguishable. Also log r2 during training: if it is literally c_j − 1 for every step, the fairness term is action-independent. If instead r2 was implemented as J − 1, the paper must be revised; the test would then be to verify that J is recomputed per candidate action and that the policy responds to c_j.","verdict_should_be":"REJECT","load_bearing_attack":"The DQN reward in Section III-A3 is r = c_d·r1 + c_j·r2, with r1 = (min_d − cur_d)/max_d and r2 = c_j − 1. For a fixed state, r2 has the same value no matter which pending request is selected; it is an additive constant. In Q-learning, adding a constant to every reward in a state shifts all Q-values by the same state-dependent amount and leaves the argmax over actions unchanged. Therefore varying c_j cannot make the policy prefer fairer request sets; only c_d scales the delay-based signal. The Table IV 'fairness-biased' result is then better explained by the reduced weight on delay minimization than by any learned fairness objective. The conclusion in Section V that the 'reward mechanism learns from the execution time' and the Abstract's 'balancing delay and fairness' are not supported by Eqs. 8–10 as printed. If r2 was intended to be J − 1 (Jain's index), the symbol c_j is overloaded and the paper does not define how J is computed during stepwise decisions; as written, the fairness coefficient is inert.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Deep Q-Network (DQN) based scheduler for entanglement requests in quantum repeater networks. The authors build a hybrid simulation framework that uses NetSquid to generate lookup tables for entanglement fidelity and execution time, and a discrete-time network simulator that queues, executes, and drops requests. The DQN is trained to select, at each step, the pending request that maximizes a reward combining a delay term and a fairness term. The scheduler is compared against Greedy, Proportional fair, and FIFO baselines under a Watts-Strogatz topology with 10 nodes and random request arrivals, reporting delay CDFs and Jain's fairness indices. The central claim is that the DQN can trade off delay and fairness, achieving lower or comparable delay with higher fairness than Greedy in a delay-biased configuration, and higher fairness than Proportional fair in a fairness-biased configuration.","tokens_in":7641,"tokens_out":3402,"duration_ms":36589,"significance":"If the claims hold, the paper makes a useful contribution by bridging quantum-level fidelity simulations (NetSquid) with network-layer scheduling simulation, and by demonstrating a learning-based scheduler that can balance QoS objectives in entanglement request scheduling. The inclusion of queuing, pending, and dropping behaviors, and the use of random topologies, are strengths. However, the central quantitative claims currently rest on an ill-defined reward formula and an incomplete evaluation methodology; the significance of the work will depend on whether those issues can be resolved.","major_comments":[{"comment":"As written, the fairness reward r2 is action-independent: Eq. (9) defines r2 = c_j − 1, which is a constant for a given coefficient c_j. In Q-learning, adding a constant to every reward in a state shifts all Q-values equally and cannot change the argmax over actions. Therefore varying c_j cannot make the DQN prefer fairer request sets, and the claimed delay-fairness trade-off is not implemented by the stated equations. The conclusion in Section V that the 'reward mechanism learns from the execution time' and the Abstract's 'balancing delay and fairness' are not supported by the equations as printed. If the intent was r2 = J − 1 for some online Jain index, the symbol c_j is overloaded and the paper does not define how J is computed during stepwise decisions. This needs to be corrected and the experiments repeated with a properly defined fairness reward.","section":"§III-A3, Eqs. (8)-(10)"},{"comment":"The reward r1 uses min_d and max_d, which the text states are 'calculated using the execution time of each request after all the requests are executed.' If the reward for a stepwise decision depends on the full future schedule, the training signal has a look-ahead bias: the agent is effectively told the outcome of the entire episode, which is not available during deployment. The authors should clarify whether rewards are computed causally at each decision step (e.g., from statistics of the current queue only) or whether this is an offline, non-causal training procedure; in the latter case, the learned policy's real-time applicability is not established.","section":"§III-A3, Eq. (8)"},{"comment":"The normalized gain is defined as (J_DQN − J_baseline)/(J_max − J_min), but J_max and J_min are never defined. Without these values, the reported gains of 'at least 14%' and '12.8%' are not interpretable. If, for example, J_max and J_min are taken as 1 and 0, the actual gains shown in Tables III and IV are roughly 2 to 4 percentage points, which is far smaller than the headline percentages. The authors must specify how J_max and J_min are chosen and report the unnormalized differences as well.","section":"§IV-B, Tables III and IV"},{"comment":"The simulation results are reported as single numbers with no error bars, confidence intervals, or number of independent runs. Delay CDFs and Jain's indices are stochastic quantities, and the observed differences between DQN and the baselines are small (e.g., 0.4893 vs. 0.4704 in Table III). Without multiple seeds or a statistical significance test, the claim that the DQN 'achieves higher fairness' than Greedy or Proportional fair is not firmly supported. The authors should repeat the simulations with several random seeds and report mean and variance.","section":"§IV-A and Tables III-IV"}],"minor_comments":[{"comment":"The Greedy and Proportional fair baselines are described only loosely; in particular, Algorithm 2 does not specify how the 'likelihood of being chosen' is computed from execution time or distance. A precise definition is needed for the comparisons to be reproducible.","section":"§III-B and §III-C"},{"comment":"The topology parameter is labeled 'Watts-Strogetz' (typo for Watts-Strogatz), and the paper does not state whether the same graph instance is used across all experiments or whether different graphs are sampled; this should be clarified.","section":"§IV-A, Table II"},{"comment":"The DQN training details are incomplete: network architecture, learning rate, replay buffer size, target network update frequency, and exploration schedule are not reported. These are needed for reproducibility, especially since the paper claims stable training with Double DQN.","section":"§III-A"},{"comment":"The assumption that requests are processed sequentially is stated but not justified, and its effect on the generality of the scheduling conclusions is not discussed. The paper would be stronger if it either justified this assumption for the target quantum network scenarios or acknowledged it as a known limitation in the conclusion.","section":"§II-C"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely topic and the simulation framework has merit, but the central reward design appears to be non-functional as written, and the evaluation lacks error bars and defined normalization constants. The issues are local enough to be fixable, but the authors will need to re-derive the fairness reward, rerun the experiments, and present the results with proper statistical support. I would also encourage the authors to release the simulation code and seeds to aid reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper builds a DQN scheduler for entanglement requests in quantum repeater networks, using NetSquid-derived lookup tables for fidelity and execution time, and compares against Greedy, Proportional fair, and FIFO. That simulation pipeline is a reasonable piece of engineering, and the problem—scheduling queued entanglement requests with conflicting delay and fairness goals—is real and underserved.\n\nThe trouble is that the reward function does not do what the authors claim. As printed, Eq. (9) is r2 = c_j − 1, which is a constant. In Q-learning, adding a state-independent constant to every reward shifts Q-values but doesn't change the argmax. So varying c_j cannot alter the policy at all. Even the delay term r1 is multiplied by c_d; scaling all action-dependent rewards by a positive constant also leaves the policy unchanged. That means the \"tunable\" trade-off in Sections III and V is not implemented as written. The fairness-biased result in Table IV is better explained by the reduced weight on r1 (or by random training variation) than by any learned fairness objective.\n\nThis is a load-bearing flaw, not a cosmetic one. The paper's main conclusion—that DQN can balance delay and fairness by adjusting reward coefficients—is unsupported. If Eq. (9) was meant to be r2 = J − 1 (Jain's index), the authors need to define J at the stepwise decision point and rerun the experiments; as submitted, the symbol c_j is overloaded and the objective is incoherent.\n\nThere are also secondary issues: the normalized gain uses Jmax and Jmin without definition, there are no error bars or seeds, only one topology is tested, and DQN hyperparameters are missing. Those are fixable. The reward bug is the one that matters.\n\nDespite this, the underlying idea is not silly. The NetSquid-to-network-simulator bridge and the scheduler comparison framework could be useful once corrected. I'd send it to peer review rather than desk reject, but with a clear request to fix the reward formulation and re-evaluate. As it stands, I wouldn't cite the empirical results.","headline":"Useful network-simulation framework, but the DQN reward as written makes the fairness coefficient inert, so the claimed delay–fairness trade-off is not implemented.","tokens_in":8136,"tokens_out":4097,"would_cite":false,"duration_ms":41735,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Deep Q-Network scheduler can balance delay and fairness for entanglement requests in quantum repeater networks, matching Greedy's low delay while improving fairness.","keywords":["entanglement request","quantum repeater network","deep Q-network","request scheduling","delay time","fairness","quantum network simulation","reinforcement learning"],"falsifier":"Re-run the medium-load simulation on the same ten-node topology but with link distances drawn from a spread of, say, tenfold; if the delay-biased DQN's fairness gain over Greedy falls below the reported normalized 14%, the homogeneous-link assumption is load-bearing.","tokens_in":7168,"feed_emoji":"⚛️","tokens_out":8810,"duration_ms":77227,"temperature":0.7,"pith_summary":"This paper argues that a Deep Q-Network (DQN) can schedule entanglement requests in quantum repeater networks more effectively than Greedy, Proportional fair, and FIFO scheduling. The authors build a network simulator that runs in time slots, queues unfinished requests, drops requests that exceed a maximum execution time, and draws fidelity and execution-time values from quantum-simulation lookup tables. They show that a single DQN architecture trained with different reward coefficients produces schedulers with different delay-fairness trade-offs. This gives quantum network operators a practical way to tune request scheduling to an application's priorities without hand-crafting heuristics.","feed_headline":"Quantum DQN scheduler matches greedy delay, beats it on fairness","feed_subtitle":"Fairness-biased DQN also tops proportional fair by 12.8% while holding delay.","key_machinery":"The central object is a Double DQN scheduler that turns request scheduling into sequential action selection: at each step it computes a Q-value for every pending request from a binary matrix encoding source and destination nodes, then serves the request with the highest value. The reward is a weighted sum of a normalized delay-reduction term, $r_1 = (\\min_d - \\mathrm{cur}_d)/\\max_d$, and a fairness term derived from the change in a fairness index; the coefficients $c_d$ and $c_j$ set the bias. This reward design is what lets one architecture express both delay-priority and fairness-priority schedulers.","core_discovery":"The scheduler picks pending entanglement requests one at a time, scoring each with a Q-value and serving the highest-scoring request until all are processed. With a delay bias (delay coefficient 0.9, fairness coefficient 0.1), the DQN's completion-delay distribution is close to Greedy's, yet its fairness index is 0.4893 in medium load and 0.5323 in low load, versus Greedy's 0.4704 and 0.5173, a normalized gain of at least 14%. With a fairness bias (delay 0.15, fairness 0.85), the DQN reaches a fairness index of 0.5927 versus Proportional fair's 0.5778, a normalized gain of 12.8%, while its delays remain similar to Proportional fair's. The authors take this as evidence that a DQN trained on simulation-derived lookup tables can traverse the delay-fairness trade-off in quantum request scheduling.","pith_inferences":["The reward decomposition into delay and fairness terms is not tied to DQN; applying the same reward with other reinforcement-learning algorithms such as policy-gradient methods would test whether the scheduler's flexibility is a property of the reward or of the specific Q-network.","Replacing the fixed-size binary state matrix with a graph-based encoder could let a single agent generalize across topologies and arrival counts, which the current approach handles by training separate DQNs.","The homogeneous-link assumption means the reported gains are likely an upper bound; on real hardware, adding link-distance and quality features to the state could recover some of the advantage.","Because the reward includes a fairness index computed over the whole request set, the method may be sensitive to the arrival process; testing with bursty or non-uniform arrivals would show how robust the trained policies are."],"forward_implications":["Network operators can pre-train DQN schedulers with different reward coefficients and pick among them as traffic or application priorities change, since the same architecture covers both delay- and fairness-biased policies.","On homogeneous ten-node topologies, the delay-biased DQN matches Greedy's delay behavior while improving fairness by a normalized margin of at least 14%.","The fairness-biased DQN improves on Proportional fair's fairness index by a normalized 12.8% at similar delay levels, showing that the approach does not simply trade all delay for fairness.","The time-slot simulator with queuing and dropping gives a testbed for other reinforcement-learning schedulers, since it captures network behaviors that pure quantum simulators do not."],"supporting_citations":[{"why":"prior deep reinforcement learning for quantum entanglement routing, the approach this work extends","marker":"[10]"},{"why":"earlier study of request scheduling in quantum networks whose trade-off analysis motivates the comparison","marker":"[11]"},{"why":"supplies the quantum simulation engine whose fidelity and execution-time distributions feed the lookup tables","marker":"[12]"},{"why":"defines fidelity, the metric used to judge whether entanglement establishment succeeds","marker":"[13]"},{"why":"gives the fidelity threshold of 0.5 used to declare an entanglement attempt successful","marker":"[14]"},{"why":"generates the random small-world topology used for training and evaluation","marker":"[15]"}],"fun_headline_variants":["DQN scheduler picks fairer entanglements, same delay as greedy","Quantum DQN scheduler improves fairness 14% without delay cost","Deep Q-Network beats proportional fair by 12.8% fairness","DQN scheduler balances delay and fairness in quantum nets","Quantum DQN scheduler: fairness up, delay flat"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training reward and the reported gains depend on the network model's assumption that all links are homogeneous (same distance and quality) and that requests are processed sequentially; a real repeater network with heterogeneous links or parallel entanglement attempts may not preserve them.","fun_headline_variants_meta":{"raw":{"variants":["DQN scheduler picks fairer entanglements, same delay as greedy","Quantum DQN scheduler improves fairness 14% without delay cost","Deep Q-Network beats proportional fair by 12.8% fairness","DQN scheduler balances delay and fairness in quantum nets","Quantum DQN scheduler: fairness up, delay flat"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001431,"raw_usage":{"total_tokens":5762,"prompt_tokens":926,"completion_tokens":4836,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":4747}},"tokens_in":542,"tokens_out":4836,"duration_ms":30825,"temperature":1.0,"reasoning_tokens":4747,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:33:42.686015+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the medium-load simulation on the same ten-node topology but with link distances drawn from a spread of, say, tenfold; if the delay-biased DQN's fairness gain over Greedy falls below the reported normalized 14%, the homogeneous-link assumption is load-bearing.","supporting_citations":[{"cited_title":"Quantum internet: networking challenges in distributed quantum computing,","cited_arxiv_id":null,"evidence_quote":"prior deep reinforcement learning for quantum entanglement routing, the approach this work extends"},{"cited_title":"Dis- tributed quantum sensing in a continuous-variable entangled network,","cited_arxiv_id":null,"evidence_quote":"earlier study of request scheduling in quantum networks whose trade-off analysis motivates the comparison"},{"cited_title":"Entanglement-based quantum communication over 144 km,","cited_arxiv_id":null,"evidence_quote":"supplies the quantum simulation engine whose fidelity and execution-time distributions feed the lookup tables"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines fidelity, the metric used to judge whether entanglement establishment succeeds"},{"cited_title":"Dqra: Deep quantum routing agent for entanglement routing in quantum networks,","cited_arxiv_id":null,"evidence_quote":"generates the random small-world topology used for training and evaluation"}],"review_version":1}