{"id":"46098e6f-2f5d-4a72-84aa-f39dcf9821d2","arxiv_id":"2509.09637","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A GNN-learned backlog plus optimal-transport scheduling reduces delay in drift-plus-penalty network routing while claiming to keep throughput guarantees.","lead":"This paper trains a graph neural network to learn the backlog function in drift-plus-penalty routing, and uses Sinkhorn optimal transport to schedule links. In simulations, the learned routiner reports lower queueing delay than backpressure, shortest-path, and queue-biased baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Penalty weight V is never used in Algorithm 1, f_phi, or the Sinkhorn objective, so the penalized DPP experiments in §VI-C do not implement a DPP trade-off.","rationale":"The reader's weakest assumption identified exactly the same gap: V is absent from the described algorithm, so the penalized experiments in Fig. 2 cannot be interpreted as a DPP trade-off. My review of the full text confirms this: Section V introduces a method that 'enabl[es] practical implementations of the DPP algorithm D(U,V) for any given backlog U and penalty weight V≥0', but V never appears in Algorithm 1's input list, in the power network f_phi(G,Q,U,S), in the scheduling LP (W), in the entropy-regularized OT problem (E), or in the stated training objectives. Therefore, as described, the algorithm's outputs are invariant to V, and the varying penalty curves in §VI-C must arise from an undocumented mechanism or are not genuine DPP trade-offs. This is a load-bearing concern because the paper's abstract and contributions explicitly claim joint optimization of power allocation and routing with a penalty, and Theorem 1 is invoked as the theoretical anchor for the penalized setting. The no-penalty delay improvements may still stand, and the theorem itself is a standard Lyapunov argument, so a conditional verdict remains appropriate: the authors must either specify how V enters the implementation or restrict the claims to the unpenalized case. No stronger rejection is warranted because the theoretical framework is sound and the missing V is a fixable gap, not a contradiction in the mathematics.","tokens_in":14420,"tokens_out":5281,"duration_ms":63737,"concrete_test":"Run the forward pass of Algorithm 1 exactly as specified in Section V with two values of V (e.g., 0.05 and 0.45) on identical seeds and all other inputs fixed. If the outputs (P,µ) and resulting penalty/queue-ratio are bit-identical, the reported V-dependence in Fig. 2 cannot be produced by the described algorithm; if they differ, trace where V enters and document the missing mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim covers joint power allocation and routing with a penalty. The DPP problem (D) in §IV includes -V p(P,S), and Theorem 1 guarantees optimal penalty as V→∞. But the described implementation omits V: Algorithm 1's inputs are (Q,S,η,G,z_U,z_P); the power network is f_phi(G,Q,U,S); the scheduling objective (W) and its OT relaxation (E) contain W+·µ and entropy but no -V p(P,S); and training is only to maximize (E). Thus for fixed (Q,S,G,z_U,z_P,η) the outputs P,µ are independent of V. Yet §VI-C reports penalty and queue-ratio curves versus V and says methods achieve 'different balance' at the same V. This is internally inconsistent: either V enters through an unspecified mechanism (e.g., as input to f_phi, or as a multiplier in the training loss), or the penalty curves are not DPP trade-offs. Since no code is provided, the experimental penalty claim is unsupported as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a learned variant of the drift-plus-penalty (DPP) method for joint transmit-power allocation and multi-hop routing. A GNN computes a neural backlog U_N and a power allocation f_phi; link scheduling is then posed as an optimal-transport problem and solved approximately with Sinkhorn iterations. Theorem 1 gives a sufficient condition for throughput optimality and asymptotic penalty optimality when the learned backlog is uniformly close to the true queue. Experiments on random geometric networks compare neural backlogs against SP, QSP, and BP baselines, reporting queue-ratio (delay) improvements under several penalties, out-of-distribution rates, and time-varying topologies.","tokens_in":14662,"tokens_out":15602,"duration_ms":169097,"significance":"If the theoretical and experimental claims hold, the paper would make a useful contribution: a differentiable, distributed DPP implementation that can be trained end-to-end while retaining stability guarantees, plus an OT-based scheduling surrogate with entropy regularization. The idea of learning a bounded backlog function is natural and the empirical gains in the no-penalty setting appear plausible. However, the manuscript as written has a load-bearing gap: the penalty weight V is never used in the described implementation, so the penalized DPP experiments are not internally consistent. The proof of Theorem 2 also contains a suspicious construction. These issues are fixable but need to be addressed before the central claims can be accepted.","major_comments":[{"comment":"Algorithm 1's inputs are (Q, S, η, G, z_U, z_P); the power network f_phi(G, Q, U, S) does not take V; the scheduling objectives (W)/(E) contain W+·μ and entropy but no -V p(P, S); and φ is trained to maximize (E). Thus, for fixed inputs, the outputs P and μ are independent of V. Yet §VI-C reports average penalty and queue ratio versus V and states that methods achieve 'different balance' at the same V. The paper never specifies a mechanism by which V enters the power network, the scheduling weights, or the training loss. As written, the penalized experiments do not implement D(U, V), so the V-curves cannot support the DPP trade-off claim of Theorem 1.","section":"§V, Algorithm 1, §VI-C"},{"comment":"Theorem 1 states throughput optimality for all V≥0 and p→p* as V→∞ without qualification. The proof in Appendix A relies on an 'optimal S-only algorithm' satisfying E[δ_i,c+λ_i,c]<0 for i≠c, an assumption stated only there. This assumption is load-bearing for the penalty part of the theorem: it supplies the θ·Q term in the drift bound. The assumption should either be stated in the theorem statement or proved from standard capacity-region conditions; otherwise the theorem overclaims.","section":"§IV-A, Theorem 1 and Appendix A"},{"comment":"The construction of a feasible solution to (T) from a solution of (W) is invalid as written. Step 2 says 'We set M_ij=0 for all elements where row i or column j is saturated.' If a row is saturated by a positive entry placed in Step 1, zeroing the entire row destroys the row sum; similarly for columns. The residual subproblem cannot then restore the original row and column targets. Since Theorem 2 is used to justify the clipping step after Sinkhorn, a correct proof is needed.","section":"Appendix B, proof of Theorem 2"}],"minor_comments":[{"comment":"The caption says 'The for-loop over V is shown for only for simplifying the exposition', but the algorithm contains a for-loop over nodes i∈V, not over the penalty weight V. This is confusing because V is also used for the penalty weight.","section":"Algorithm 1 caption"},{"comment":"The bound B in Theorem 1 is called A=10 in the experiments without comment. Also, the sigmoid output layer for Neural-B is said to 'satisfy the bound ||Q−U_N||∞≤B', but the exact construction that guarantees this bound is not given.","section":"§IV-B3 and §VI"},{"comment":"The neural backlogs are trained to minimize ||Q||_1, which is also the basis of the queue-ratio evaluation metric. This makes the no-penalty comparisons partly a fitting-to-the-target exercise. The out-of-distribution and time-varying results help, but the paper should acknowledge this circularity and, if possible, report a metric not directly optimized.","section":"§IV-B3 and §VI"},{"comment":"No code or detailed hyperparameters are provided. Given the V-omission issue, the experimental penalty results are not reproducible from the text alone. Please add an implementation appendix or release code.","section":"§VI-A"}],"recommendation":"major_revision","confidential_remarks":"The V-omission is serious: either the experiments were run with an additional input not described in the paper, or the penalty curves are not DPP trade-offs. I would strongly encourage requesting code or a precise description of how V enters f_phi or the training objective during revision. The S-only assumption in Theorem 1 should also be made explicit, and the proof of Theorem 2 needs a careful rewrite."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the no-penalty results look credible: the learned backlog plus Sinkhorn scheduler consistently lowers queue ratios across rates, and the out-of-distribution and mobility experiments give some independent grounding. That part is worth taking seriously. Second, the paper's central theoretical claim about the penalty weight V is undermined by the implementation. V is not in Algorithm 1, not in the power network f_phi(G,Q,U,S), and not in the Sinkhorn objective (E). The training objective for phi is to maximize (E), which contains W+·mu and entropy but no -V p(P,S). Yet Section VI-C plots average penalty and queue ratio against V and claims different balances at the same V. Unless the authors have an undocumented mechanism, those curves do not represent a DPP trade-off; they are just the same outputs relabeled. The stress-test note is accurate, and it is a load-bearing flaw for the penalized-experiment claim.\n\nWhat is genuinely new and good here: the optimal-transport reformulation of the scheduling subproblem (Theorem 2) is a clean and practical trick, and the sparse Sinkhorn implementation is a real contribution that makes the whole pipeline differentiable and GPU-friendly. The neural backlog trained via direct queue-size minimization is a reasonable extension of prior GNN backpressure work, and the paper is honest about the gap between stability guarantees and empirical performance. Theorem 1 is a standard Lyapunov argument and is correct under the acknowledged S-only slack assumption; the proof is fine.\n\nSoft spots, in proportion. The V problem is severe and must be addressed before the joint power-and-routing claims can be trusted. The abstract says \"consistent improvements over common baselines\" and mentions \"existing ML-based approaches,\" but there is no comparison to the GNN backpressure baseline [41] that is cited. No code, data, or training hyperparameters are provided, so the results are not reproducible as-is. These are fixable, but they are real.\n\nWho is this for: networking researchers working on learned backpressure and resource allocation. It deserves a serious referee—the Sinkhorn scheduling idea alone warrants attention—but only after the V mechanism is clarified and code is released. I would send it to peer review with a clear request for major revision: either show exactly how V enters the algorithm and training, or restrict the claims to the no-penalty setting and compare fairly with [41].","headline":"A promising no-penalty result and a clever Sinkhorn scheduling trick, but the penalty weight V never enters the described algorithm, so the paper's main DPP trade-off claim is not supported as written.","tokens_in":15169,"tokens_out":2592,"would_cite":false,"duration_ms":31979,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that replacing the back-pressure backlog with a learned, bounded graph-neural-network backlog and scheduling with Sinkhorn optimal transport preserves drift-plus-penalty optimality while consistently reducing queue ratio (d","keywords":["drift-plus-penalty","back-pressure routing","graph neural networks","Sinkhorn algorithm","optimal transport","power allocation","Lyapunov optimization","queue stability"],"falsifier":"Run Algorithm 1 with two different penalty weights V (e.g., V=0.05 and V=0.4) on identical inputs and check whether the output power and schedule are identical; if the forward pass is invariant to V, then the claimed DPP penalty trade-off is not implemented and the Section VI-C curves cannot be attributed to the drift-plus-penalty mechanism.","tokens_in":14290,"feed_emoji":"📡","tokens_out":4180,"duration_ms":46565,"temperature":0.7,"pith_summary":"The paper aims to show that the classical drift-plus-penalty method for network routing can be made both safer and faster by learning its backlog function with a graph neural network and by solving its scheduling subproblem with an entropy-regularized optimal-transport routine (Sinkhorn). Its core theoretical result is that any backlog function within a bounded distance of the true queue lengths is throughput-optimal and approaches the optimal penalty as the penalty weight grows, so a bounded neural backlog keeps the guarantees. Empirically, the learned backlog and Sinkhorn scheduler consistently improve the queue ratio (a proxy for delay) over back-pressure, shortest-path, and QSP baselines under no-penalty, power-consumption, and energy-efficiency objectives. This matters because drift-plus-penalty methods are known to have poor delay under light load, and the paper shows the trade-off can be relaxed without sacrificing the Lyapunov-style guarantees.","feed_headline":"Learned backlogs keep drift-plus-penalty guarantees and cut delay","feed_subtitle":"A bounded neural backlog plus optimal-transport scheduling beats classical back-pressure on queue ratio across rates and penalties.","key_machinery":"The load-bearing objects are (i) the backlog function U, which replaces the raw queue difference in the drift-plus-penalty weight W = U_i,c - U_j,c and thereby controls route selection; (ii) Theorem 1, a sufficient condition linking a uniform bound on U-Q to throughput optimality and penalty optimality; and (iii) the Sinkhorn-based solver for the entropy-regularized optimal transport formulation of the scheduling LP, which computes a differentiable approximate max-weight schedule. The backlog bound is what converts a learned heuristic into a provably stabilizing controller; the Sinkhorn entropy term is what makes the schedule differentiable and spreads traffic across paths.","core_discovery":"The central claim is that drift-plus-penalty optimality survives a learned backlog. Theorem 1 states that if the backlog function U(t) stays within a uniform bound B of the actual queue vector Q(t) at every time, then the DPP algorithm is throughput-optimal for every penalty weight V and drives the expected penalty to the optimum p* as V goes to infinity. The paper then constructs two neural backlogs from a GNN: Neural-B with a sigmoid output that enforces the bound, and Neural with a linear output that does not. Scheduling is posed as a discrete optimal transport problem, solved approximately with Sinkhorn iterations, which is fully differentiable and restores gradients for training. Experi","pith_inferences":["The paper's finding that the unbounded Neural backlog often outperforms the bounded Neural-B suggests the uniform-bound condition may be sufficient but not necessary; a relaxed bound might be enough in practice, though the guarantees would then need a separate argument.","If the penalty weight V truly does not enter the forward pass as written in Algorithm 1, the penalized curves in Section VI-C should be reread as comparing methods at the same nominal V rather than as demonstrations of the DPP trade-off; a corrected implementation would pass V into the power network or scheduling objective.","The entropy-regularized schedule could be viewed as an implicit load-balancing mechanism; this suggests a testable extension where eta is annealed or learned per-node rather than fixed.","The sparse, batched Sinkhorn implementation points to a practical GPU deployment path for distributed routing, and the same architectural pattern could be applied to other LP-based network controllers."],"forward_implications":["Any bounded backlog function, including a learned one, inherits the throughput-optimality and asymptotic penalty-optimality guarantees of drift-plus-penalty control.","The Sinkhorn-based scheduler provides a differentiable surrogate for max-weight scheduling, enabling end-to-end unsupervised training of the backlog from a network model.","At low data rates, where classical back-pressure suffers random-walk delays, Sinkhorn scheduling with neural backlogs yields the largest queue-ratio gains.","The method extends naturally to time-varying network topologies and out-of-distribution arrival rates without retraining per topology.","The entropy regularization parameter eta acts as an additional congestion-control knob: lower eta spreads traffic and improves weak backlogs like BP and QSP."],"fun_headline_variants":["Neural DPP cuts delay, keeps optimality","Learned backlogs lower delay, same guarantees","Optimal transport makes DPP leaner","AI routing: less delay, same throughput","Bounded neural backlog beats back-pressure"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper's theoretical and experimental claims hold only if the implemented algorithm actually incorporates the penalty weight V that the drift-plus-penalty objective names; as written, Algorithm 1 never receives V, so the penalty trade-off curves rest on an unspecified mechanism for V entering the power or scheduling choices.","fun_headline_variants_meta":{"raw":{"variants":["Neural DPP cuts delay, keeps optimality","Learned backlogs lower delay, same guarantees","Optimal transport makes DPP leaner","AI routing: less delay, same throughput","Bounded neural backlog beats back-pressure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000137,"raw_usage":{"total_tokens":953,"prompt_tokens":679,"completion_tokens":274,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":423,"completion_tokens_details":{"reasoning_tokens":206}},"tokens_in":423,"tokens_out":274,"duration_ms":4133,"temperature":1.0,"reasoning_tokens":206,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T18:45:06.456354+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 with two different penalty weights V (e.g., V=0.05 and V=0.4) on identical inputs and check whether the output power and schedule are identical; if the forward pass is invariant to V, then the claimed DPP penalty trade-off is not implemented and the Section VI-C curves cannot be attributed to the drift-plus-penalty mechanism.","supporting_citations":[],"review_version":1}