{"id":"6b261e58-2e9e-4e3b-ac7a-ca7bb1395f49","arxiv_id":"2604.05175","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A U-GNN diffusion policy trained on primal-dual expert samples generates near-optimal, near-feasible stochastic power allocations for ergodic wireless networks and transfers across QoS and size.","lead":"This paper trains a graph-based diffusion model to instantly produce randomized power allocations that match a slow optimization algorithm for wireless networks. If the results hold, fast sampling could replace expensive online optimization in dense wireless systems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Near-optimality claim depends on an unverified expert: the primal-dual guarantee (11)-(12) has unstated assumptions and an unquantified parametrization gap (footnote 1), and the K=200 late-iterate window is never checked against an external optimum.","rationale":"I read the paper in good faith. The mathematical framing (1)-(13) is coherent: if bDx is near-optimal and the diffusion model matches it, time-sharing samples should amortize the expert. The strongest independent support is the existence of an asymptotic guarantee (11)-(12), but it is cited from author-overlapping work, not restated with assumptions, and footnote 1 explicitly says the parameterized version has an additional expressiveness gap. Because the expert uses a shallow GNN for the inner Lagrangian maximization and only K=200 iterates are retained, the target distribution may not be close to D*_x. The per-iterate infeasibility in Fig. 1 is not itself a flaw—feasibility is in the ergodic average—but it makes checking the actual averaged slack essential. The empirical section compares against the expert itself and against simple deterministic baselines (AP/FP); neither establishes near-optimality in an absolute sense. I do not claim the results are wrong; I claim the central claim is currently unsupported at its foundation. The missing code/data and external baseline are addressable, so a conditional verdict is appropriate. Since the reader already assigned CONDITIONAL, my recommendation is UNCHANGED (same conditional status).","tokens_in":8537,"tokens_out":6479,"duration_ms":64825,"concrete_test":"Run the same primal-dual expert on a small network (e.g., N=2 or 3) from the paper's channel model, using the same 3-layer GNN parametrization, early stopping, and K=200 late-iterate window. Compare the time-averaged utility and min-rate feasibility of that window against a truly optimal stochastic policy obtained by fine exhaustive discretization of the power allocation or by an independent convexified time-sharing solver. If the expert window's utility falls below the true optimum by more than the reported margins (or if its constraint slack violates the p1/p5 thresholds), then the expert target bDx is not near-optimal and the learned policy's central claim is void. This check directly tests the unstated assumptions behind (11)-(12) and the footnote-1 parametrization gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that time-sharing U-GNN diffusion samples achieves near-optimal ergodic sum-rate and near-feasible min-rate constraints. This can only hold if the expert distribution bDx in (13) is itself near-optimal and feasible, because the diffusion model is trained to imitate bDx and the paper gives no bound on the imitation-to-utility gap. The sole support for bDx is the cited theorem (11)-(12), whose assumptions are not stated. Moreover, the paper's footnote 1 admits that the parameterized version carries an additional optimality gap depending on the expressiveness of the parametrization Phi; since the expert's primal maximization (9) is implemented with a shallow 3-layer GNN, this gap is unquantified and could exceed the reported margins. The practical bDx is only the final K=200 iterates of a 20,000-iteration run with early stopping, but no burn-in K0 selection is reported and the window is not verified to satisfy the time-average guarantee. Fig. 1 even shows at least 20% of receivers infeasible at each iterate, so the feasibility claim rests entirely on averaging over the window. Fig. 2 compares the learned policy only to this expert, not to an external near-optimal baseline, so the headline 'near-optimal' is not independently checked. Footnote 1 and the deferred code release are explicit admissions that this load-bearing support is missing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a conditional diffusion-model policy for constrained ergodic wireless resource allocation. It uses a primal-dual algorithm to generate expert allocation trajectories, collects the final K=200 primal iterates as an empirical expert distribution, and trains a U-Net-style graph neural network (U-GNN) denoiser to imitate this distribution. At inference, DDIM sampling produces allocations that are time-shared over slots. In a 400-user power-control case study, the paper reports that the learned policy closely matches the expert in tail-rate percentiles and mean-rate utility, while generalizing across QoS levels and transferring to larger networks.","tokens_in":8897,"tokens_out":5028,"duration_ms":52418,"significance":"The topic is timely, and the graph-signal parameterization with a U-GNN denoiser is a sensible way to obtain permutation equivariance and size transferability. The paper also correctly emphasizes that stochastic time-sharing policies are needed for constrained ergodic problems, rather than deterministic allocations. If the near-optimality claims are verified, the approach would be a valuable amortized alternative to online primal-dual methods. However, the current evidence for near-optimality is indirect: the learned policy is compared only against the expert that generated its training data, and the expert's own near-optimality is inherited from a cited theorem with unstated assumptions. The paper nonetheless makes a useful empirical contribution and the gaps appear addressable with additional analysis and experiments.","major_comments":[{"comment":"The expert distribution bDx is the sole ground truth for training and for the headline comparisons, but its near-optimality is not established within this manuscript. Equations (11)–(12) are cited from [3] without stating the assumptions on f0, f, X, or the step-size condition; footnote 1 admits that the parameterized version has an additional optimality gap 'depending on the expressiveness of the parametrization Φ.' Since the expert's primal maximization (9) is implemented by a shallow 3-layer GNN, this gap is unquantified and could exceed the reported margins. Please state the theorem hypotheses, verify them for the power-control problem, or provide a numerical certificate of expert optimality (e.g., an upper bound from a convex relaxation or comparison with a global solver on small instances).","section":"§2.2, Eqs. (11)–(13), footnote 1"},{"comment":"The construction of bDx uses the final K=200 primal iterates, but the burn-in K0 is never reported and its choice is not justified. Fig. 1 shows a long transient and per-iteration constraint violations for at least 20% of receivers at all iterations, so feasibility rests entirely on averaging over the selected window. In addition, the 200 iterates are serially correlated; treating them as i.i.d. samples in the diffusion training objective (4) may bias the denoiser. Please report K0, give the time-averaged utility and constraint values over the chosen window, and clarify whether the samples are sub-sampled or how the correlation is otherwise accounted for.","section":"§3.1, Eq. (13), Fig. 1"},{"comment":"The central claim is that time-sharing U-GNN samples is 'near-optimal,' but the only optimality benchmark is the expert policy, which is also the training target. The AP baseline is the conditional mean of the expert and FP is full power; neither provides an independent upper bound or optimality certificate. Thus the near-optimality claim is circular unless the expert is independently certified (see comment above). At minimum, include a comparison to a known upper bound or to exact/exhaustive solutions on a small network, and report variance over random seeds and test networks; Fig. 2 currently reports a single trajectory without error bars.","section":"§3.2, Fig. 2"}],"minor_comments":[{"comment":"The conditional KL divergence expression is typeset ambiguously; the argument of DKL should be written with an explicit integral over x to avoid confusion about the conditioning on H.","section":"§2, Eq. (2)"},{"comment":"The noise term W N0 in the denominator is dimensionally correct if N0 is a power spectral density, but this should be stated explicitly for clarity.","section":"§3, Eq. (14)"},{"comment":"The caption refers to a 'golden shaded region' marking the convergence regime, but no such region is visible in the figure; please clarify or redraw. Also define K0 in the caption.","section":"Fig. 1 caption"},{"comment":"The percentiles p1, p5, p10 are used without a precise definition; state whether they are percentiles over receivers of per-receiver ergodic rates. Also, the generalization and transferability claims are only qualitative; adding numerical tables would strengthen the paper.","section":"§3.2, Figs. 3–4"},{"comment":"Reference [3] contains a typo: 'Transactiosn' should be 'Transactions.'","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on the authors' own prior work (refs [3], [11], [12], [14]) for the expert algorithm and backbone architecture. This is not disqualifying, but the central guarantee comes from [3] and is not restated or verified here, which is a fairness concern for a standalone paper. The empirical setup is promising and the requested additions (state the theorem assumptions, verify expert optimality, add non-circular benchmarks, report K0 and variability) are within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a reasonable incremental step: it takes the authors' own diffusion-for-resource-allocation framework and applies it to stochastic ergodic power control with a U-GNN denoiser, showing that a single DDIM pass roughly tracks a primal-dual expert on tail rates and mean utility, with some generalization across QoS levels and network sizes. That is a legitimate contribution, though not a new idea—the framework is in [12] and the U-GNN backbone in [11].\n\nWhat it does well: the problem formulation (1) is clean, and the point that deterministic policies are provably suboptimal for ergodic constraints is worth making. Using primal-dual late iterates as training samples for a conditional generative model is natural and sensible. The deterministic baselines failing the tail constraints illustrate the value of randomization. The figures, if reproducible, are convincing that the learned policy imitates the expert well.\n\nThe soft spots are where the load-bearing claim rests. The near-optimality of the expert is not established in this paper: equations (11)-(12) are cited from [3] with no assumptions stated, and footnote 1 admits an additional gap from the GNN parametrization. Since the expert is itself a 3-layer GNN, that gap is unquantified and could be large. The K=200 sample window is used without discussing burn-in or checking against an independent optimum; Fig. 1 shows 20% of receivers infeasible per iteration, so the feasibility claim is entirely about time averages. The evaluation also lacks error bars and seeds, code is deferred, and the only stochastic baseline is the very expert the model was trained to imitate—which measures imitation fidelity, not optimality. None of this is fatal; it is fixable. State the assumptions, quantify or empirically bound the parametrization gap, add an external near-optimal baseline, and release code and data.\n\nVerdict: send it out. It is a serious, coherent piece of engineering that a competent referee can engage with. The central idea is sound; the evidence needs to be stronger before the near-optimal claim can be taken at face value. I would not desk-reject it.\n\nFor your reading group: maybe, if someone cares about diffusion models for wireless optimization.","headline":"A sensible, incremental extension of the authors' own diffusion-for-resource-allocation line; the empirical story is plausible but the near-optimality claim leans on an unproven expert and an unquantified GNN gap.","tokens_in":9424,"tokens_out":4098,"would_cite":false,"duration_ms":42920,"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":"A graph-neural-network diffusion model can learn near-optimal stochastic power-control policies, so a single sampling pass replaces thousands of primal-dual iterations.","keywords":["wireless resource allocation","diffusion models","graph neural networks","primal-dual algorithms","power control","ergodic optimization","stochastic policies","graph signal generation"],"falsifier":"Take a small network (e.g., N=2 or N=3) where the true optimal stochastic policy of problem (1) can be computed exactly by discretization or exhaustive search, train the same U-GNN diffusion model on the primal-dual expert dataset, and compare the time-shared ergodic utility and constraint feasibility of the learned policy against the true optimum; if the gap exceeds the primal-dual suboptimality bound O(ηλ) plus a small numerical tolerance, the paper's near-optimality claim is falsified.","tokens_in":8360,"feed_emoji":"📡","tokens_out":3561,"duration_ms":32466,"temperature":0.7,"pith_summary":"The paper tries to show that a diffusion model built from graph neural network blocks can learn the stochastic resource-allocation policies that solve constrained ergodic wireless optimization problems. Because the optimal policy is a distribution over power allocations—randomized time-sharing is needed to meet rate constraints—the paper treats allocations as graph signals and trains the model to imitate an expert conditional distribution produced by a primal-dual algorithm. At test time, a single denoising sampling pass replaces thousands of online primal-dual iterations. In a 400-user power control study, the paper reports that time-averaging the generated allocations achieves near-optimal ergodic sum-rate and near-feasible minimum-rate guarantees, and that the learned policy generalizes to unseen QoS levels and larger networks.","feed_headline":"Diffusion model learns near-optimal power allocations in one pass","feed_subtitle":"Time-sharing the generated allocations nearly meets rate constraints and generalizes to larger networks.","key_machinery":"The key object is a conditional diffusion model whose denoiser is a U-GNN: a U-shaped cascade of graph neural network blocks with skip connections, conditioned on the channel-state graph and a three-feature node signal (direct-link strength, aggregate interference at full power, QoS requirement). Training uses the DDPM noise-prediction objective over clean samples drawn from the expert empirical distribution; inference uses the DDIM reverse process. The mechanism works by learning to generate allocations as graph signals, so that the learned conditional distribution can be sampled in one pass and time-shared to realize the ergodic optimum.","core_discovery":"The central claim is that the conditional distribution over resource allocations induced by a U-Net-style graph neural network diffusion model, trained to match the late-iterate empirical distribution of a primal-dual expert, is close enough to the optimal stochastic policy that time-sharing its samples yields near-optimal ergodic utility and near-feasible constraints. The paper establishes this by constructing expert datasets from the last K=200 primal-dual iterates for 128 network configurations, training the diffusion denoiser with the standard DDPM objective, and then evaluating the learned policy through DDIM sampling with 100 steps. The reported comparisons show the learned policy trac","pith_inferences":["Beyond the paper: the learned diffusion model could be used as a warm start or proposal for the primal-dual algorithm, potentially reducing the number of online iterations while retaining optimality guarantees.","Beyond the paper: the quality of the learned policy is bounded by the quality of the expert conditionals; if the expert's convergence is slow (as Fig. 1's long transient indicates), the same diffusion training procedure could be applied to a faster or more sample-efficient expert to improve tail performance.","Beyond the paper: the DDIM stochasticity parameter σ_k offers a principled way to tune the exploration-exploitation trade-off in the generated policy, a control lever the paper does not explore.","Beyond the paper: the reported near-feasibility gap at the p1 tail (0.45 vs 0.6 bits/s/Hz) suggests the learned policy may not perfectly preserve the expert's ergodic feasibility for the most constrained receivers; a direct comparison of cumulative distribution functions of generated versus expert allocations would make this gap explicit."],"forward_implications":["If the claim holds, wireless resource allocation can be amortized: a single forward pass of a trained diffusion model replaces thousands of primal-dual iterations at inference.","Randomized (time-sharing) policies become practically accessible, since the diffusion model outputs samples from the stochastic optimal policy rather than a single deterministic allocation.","The U-GNN structure implies scalability and transferability: policies trained on 400-user networks can be evaluated on larger networks and varied densities without retraining, as shown in the paper's size-transfer experiments.","The same conditional-generative formulation applies to other constrained ergodic problems beyond power control, such as MIMO or broadcast channel resource allocation, since the abstract problem (1) encompasses them."],"fun_headline_variants":["Diffusion model nails power control in a single shot","One-pass diffusion sampling for near-optimal wireless power","Graph diffusion learns instant allocation policies","Near-optimal power allocation via one-step diffusion","Amortized power allocation by diffusion sampling"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the empirical distribution built from the last K=200 primal-dual iterates (Eq. 13) is a faithful proxy for the true optimal stochastic policy D*_x(·|H); the paper's only support is an asymptotic guarantee whose assumptions are not stated, and footnote 1 concedes the gap 'additionally depending on the expressiveness of the parametrization Φ,' while Fig. 1 shows the per-iteration iterates remain infeasible for at least 20% of receivers.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model nails power control in a single shot","One-pass diffusion sampling for near-optimal wireless power","Graph diffusion learns instant allocation policies","Near-optimal power allocation via one-step diffusion","Amortized power allocation by diffusion sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000602,"raw_usage":{"total_tokens":2603,"prompt_tokens":658,"completion_tokens":1945,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":402,"completion_tokens_details":{"reasoning_tokens":1876}},"tokens_in":402,"tokens_out":1945,"duration_ms":14630,"temperature":1.0,"reasoning_tokens":1876,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T16:42:04.280248+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small network (e.g., N=2 or N=3) where the true optimal stochastic policy of problem (1) can be computed exactly by discretization or exhaustive search, train the same U-GNN diffusion model on the primal-dual expert dataset, and compare the time-shared ergodic utility and constraint feasibility of the learned policy against the true optimum; if the gap exceeds the primal-dual suboptimality bound O(ηλ) plus a small numerical tolerance, the paper's near-optimality claim is falsified.","supporting_citations":[],"review_version":2}