{"id":"dea4e5a9-d7ac-4d52-9176-43d230d4f283","arxiv_id":"2512.24659","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A hierarchical Stackelberg-game solver with matching, diffusion-enhanced TD3, and a KKT-based allocation rule reduces simulated task delay by 2.5% and energy by 3.1% in an IRS-aided vehicle MEC network.","lead":"This paper designs a system in which a drone-based edge server and a ground base station cooperate to run vehicle tasks, while reflecting surfaces on buildings and on the drone strengthen blocked wireless links. The authors test a hierarchical algorithm that splits decisions between servers and vehicles, reporting 2.5% lower delay and 3.1% lower energy than baseline methods in simulation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 2.5%/3.1% headline gains rest on an unseeded, unreleased simulator with unreported scalarization weights and energy-model constants; the improvement may be within run-to-run noise.","rationale":"The reader's conditional verdict is based on the same observation: the paper's central numerical claim is simulation-only, with unreported weights, no seeded statistics, and no code. My stress-test does not find a deeper mathematical contradiction; the KKT closed-form and matching mechanism are internally reasonable. The decisive weakness is that the reported 2.5%/3.1% improvements are small, and with no confidence intervals or full configuration, the comparison could easily be a product of stochastic variation or weight choices. This supports the reader's CONDITIONAL verdict rather than changing it: the authors should supply code, full parameters, multi-seed results, and violation metrics before the gain is accepted as significant.","tokens_in":28748,"tokens_out":9525,"duration_ms":109253,"concrete_test":"Release code and the full configuration (ω_i, ω_s, ω_c, r_bv, r_dv, η1–η4, U_tip, path-loss exponents), then run HOOA and the best baseline over the default scenario with at least 20 independent seeds, reporting mean±95% CI for average delay, average energy, and deadline-violation rate. If the 2.5%/3.1% gaps fall within the CI overlap, or the ordering flips under any reasonable weight choice, the headline claim is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—2.5% delay and 3.1% energy improvement—is not yet established because every reported point comes from the authors' closed simulator without multi-seed statistics, released code, or a fully specified configuration. Section 5.2 defines the scalarized costs C_i(n) and C_s(n) via weights ω_i and ω_s, and Eq. (41) uses ω_c and penalties r_bv(n), r_dv(n); none of these weights or penalty coefficients appear in Table 1. The UAV flight-energy model in Eq. (17) depends on constants η1–η4 and U_tip, also absent. Since delay and energy are conflicting objectives, any reported aggregate delay/energy value is conditional on an unreported scalarization, and a different weighting could reverse the ordering. Moreover, the original MOOP's hard deadline constraint (19d) is represented only by a penalty in the DRL reward, and no deadline-violation rate is reported; if the gain is achieved by allowing more violations, the average-delay improvement is misleading. The headline differences are small (2.5%, 3.1%); without confidence intervals, they may be within the variance induced by Gauss-Markov mobility, Rician fading, and diffusion sampling noise. Thus the load-bearing premise—that HOOA 'significantly improves' over the best baselines—is supported only by an unverifiable simulator.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an IRS-enabled low-altitude MEC architecture for vehicular networks, with an aerial MEC server on a UAV cooperating with a terrestrial MEC server and a hybrid deployment of building-installed and UAV-carried IRSs. It formulates a multi-objective optimization problem (MOOP) minimizing total task completion delay and energy consumption by jointly optimizing task offloading, UAV trajectory, IRS phase shifts, and computation resource allocation. The proposed HOOA solves the MOOP by reformulating it as a Stackelberg game: vehicles act as followers using a many-to-one matching mechanism for offloading decisions, while MEC servers act as leaders using a generative diffusion model-enhanced TD3 (GDMTD3) for trajectory and phase-shift control, plus a KKT-based closed-form computation-resource allocation. Simulations are reported against several heuristic and DRL baselines, claiming 2.5% average delay reduction and 3.1% average energy reduction, with convergence, robustness, and scalability results.","tokens_in":29173,"tokens_out":4210,"duration_ms":45555,"significance":"If the reported performance holds, the paper makes a useful systems contribution: the hybrid IRS architecture is a reasonable extension of existing UAV-MEC work, and the HOOA decomposition is an attractive way to handle mixed-integer, high-dimensional, time-coupled decisions. The internal mathematical parts are mostly clean: the convexity of the resource-allocation subproblem is proved correctly, and the closed-form KKT solution in Theorem 2 is a genuine optimal allocation for that subproblem. The algorithm is coherent and the benchmark comparisons are appropriate in structure. The main weakness is that the central quantitative claim, which is the basis for the paper's significance, rests entirely on an under-specified and unreleased simulator. Several scalarization weights, penalty coefficients, and energy-model constants are not reported, and all results are point estimates without seeding, confidence intervals, or statistical tests. Given the small size of the claimed gains, the performance advantage is not yet established.","major_comments":[{"comment":"The overall reward and the vehicle/server costs are defined through weights that are never specified. C_i(n) in Eq. (22) depends on ω_i, C_s(n) in Eq. (25) on ω_s, and the reward in Eq. (41) on ω_c and the penalties r_bv(n), r_dv(n). None of these values appear in Table 1. The UAV flight energy model in Eq. (17) also depends on η1–η4 and U_tip, which are likewise absent. Because delay and energy are conflicting objectives, all reported delay/energy numbers are conditional on this unreported scalarization, and a different weighting could alter or reverse the comparisons. Please provide the full configuration, including all weights and penalties, and ideally release code or a detailed seed/randomization protocol.","section":"§5.2.3, Eq. (41); Table 1"},{"comment":"The headline gains—2.5% average delay and 3.1% average energy—are small and are reported as deterministic point estimates. No number of independent runs, standard deviations, confidence intervals, or significance tests are given. Under Gauss-Markov mobility, Rician fading, and diffusion sampling noise, these differences may be within run-to-run variance. Moreover, the hard deadline constraint (19d) is enforced only through the penalty r_dv(n) in Eq. (41), yet no deadline-violation rate is reported. If HOOA achieves lower average delay at the cost of more deadline misses, the average-delay improvement is misleading. Please provide multi-seed statistics and the deadline-violation rate for all compared schemes.","section":"§6.2.1 and Abstract"},{"comment":"The Stackelberg reformulation and the many-to-one matching mechanism are heuristic. No theorem or argument shows that the matching output is the follower-level optimum, that the leader-follower iteration approximates a solution of the original MOOP (19), or that the resulting joint policy has any bounded suboptimality. The matching preferences in Eqs. (29)–(30) capture only transmission-related quantities and computation energy, not the full vehicle/server costs or deadline constraints; tasks left unmatched are processed locally by default. This is a plausible heuristic design, but the paper should state this limitation explicitly and, ideally, provide an optimality-gap study or an ablation showing the effect of the decomposition on the scalarized objective.","section":"§5.2.3 and §5.3.3"},{"comment":"The convergence and sensitivity analyses are qualitative. The text claims 'higher and smoother rewards' and 'more stable delay and energy trends' without quantifying variance or convergence rate, and the hyper-parameter comparisons in Fig. 5 report single training curves without repeated-seed variability. A quantitative comparison (e.g., mean final reward with error bars, or a Wilcoxon test across seeds) is needed to support the claim that HOOA has 'superior convergence stability' and that the adopted hyper-parameters are statistically justified.","section":"§6.2.2–§6.2.3"}],"minor_comments":[{"comment":"The cross-reference to the architecture figure appears as 'Fig. 3.1.1' in the text but should be 'Fig. 1'.","section":"Fig. 1"},{"comment":"The y-axis labels in Fig. 4(b), Fig. 4(c), and all panels of Fig. 5 are 'Reward', but the plotted quantities are average task completion delay, average energy consumption, and reward, respectively. The labels should match the plotted metric.","section":"Figs. 4 and 5"},{"comment":"The notation for the noise schedule is confusing: β_t is defined in terms of β_min and β_max, but the values of β_min/β_max are not given, and α_t / ̅α_t are not defined in Table 1. Please specify these values.","section":"Eq. (46)"},{"comment":"The decision process is called a POMDP, but the state space s(n) appears to include the full system information plus history. Clarify whether there is actual partial observability or whether this is an MDP with an augmented state.","section":"§5.4.1"},{"comment":"The complexity of the matching mechanism is stated as O(|I| log |I|), but Algorithm 1 has an iterative rejection loop that could in principle run more than one pass. Please clarify whether the stated complexity is an average-case or worst-case bound.","section":"§5.5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the algorithmic core is solid, but the experimental evidence for the headline claims is not yet reproducible or statistically grounded. The missing weights, penalties, and aerodynamic constants are easily fixable in a revision, so I do not see a reason for outright rejection. However, if the authors cannot provide multi-seed statistics and the missing configuration, the contribution would fall to a purely algorithmic proposal without demonstrated advantage over existing DRL baselines."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper assembles a genuinely new architecture—building-installed plus UAV-carried IRSs serving a two-tier MEC system—and solves it with a hierarchical Stackelberg decomposition, matching for offloading, diffusion-enhanced TD3 for continuous control, and a KKT closed-form for resource allocation. The KKT part (Theorem 2) is clean and correct; the convexity argument holds. The complexity analysis is careful. That part is worth building on.\n\nNow the soft spots. The central quantitative claim—2.5% delay and 3.1% energy improvements—comes from a simulator that is not released, described in Section 6 with parameters in Table 1, but Table 1 omits the weighting factors ω_i, ω_s, ω_c and the penalties r_bv(n), r_dv(n) that define the reward (Eq. 41) and the vehicle/server costs (Eqs. 22, 25). Without these, the exact objective being optimized is not specified, and because delay and energy are conflicting, different weight choices could reorder the results. The rotary-wing energy model (Eq. 17) also depends on constants η1–η4 and U_tip that never appear in Table 1. No multi-seed statistics or error bars are reported; with small gains, the difference may be within run-to-run noise. The original hard deadline constraint (19d) is only a penalty in the DRL reward, and no deadline-violation rate is reported, so a reader cannot tell whether the lower average delay comes at the cost of more violations.\n\nThe Stackelberg/matching decomposition is heuristic. It is a plausible way to break up an NP-hard MINLP, but there is no equivalence or optimality gap analysis versus the original MOOP. The matching mechanism is standard request-admit with preference lists; it produces feasible matches but no stability or optimality proof is supplied. Also, GDMTD3 appears very close to the authors' earlier diffusion-enhanced DRL in [59]; the paper does not state what is new beyond the application.\n\nThese are all fixable rather than fatal. The architecture and the KKT subproblem are worth publishing. For a serious referee, the authors must release code and exact configuration, add multi-seed statistics, fill in all weights and constants, report deadline-violation rates, clarify the gap between the MOOP and the Stackelberg game, and compare directly against [59]. With those changes this could be a solid systems paper. As it stands, treat the headline numbers as indicative, not established.","headline":"The hybrid-IRS architecture and the KKT closed-form are solid; the headline gains are plausible but under-supported by a closed simulator with unreported weights and no variance.","tokens_in":29633,"tokens_out":3150,"would_cite":false,"duration_ms":31477,"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":"Jointly optimizing offloading, UAV trajectory, IRS phases, and resource allocation in an IRS-enabled low-altitude MEC system can be solved online via a Stackelberg decomposition, cutting simulated task delay by 2.5% and energy by 3.1%.","keywords":["intelligent reflecting surface","low-altitude MEC","vehicular networks","unmanned aerial vehicle","task offloading","deep reinforcement learning","Stackelberg game","multi-objective optimization"],"falsifier":"In a field test or a channel-accurate simulator, run the learned IRS phase-shift policy and then randomly permute the phase shifts while keeping the UAV trajectory and offloading decisions identical; if delay and energy do not measurably worsen, the IRS contribution is overstated. Alternatively, add a realistic phase-configuration overhead to the delay and energy objective—if that overhead exceeds the reported 2.5% and 3.1% margins, the claimed improvement disappears.","tokens_in":28631,"feed_emoji":"🚁","tokens_out":6585,"duration_ms":60624,"temperature":0.7,"pith_summary":"The paper proposes an architecture where an aerial MEC server on a UAV and a terrestrial MEC server at a base station jointly serve vehicles, while building-mounted and UAV-carried IRSs reflect signals around urban blockages. It formulates a multi-objective optimization problem that minimizes task completion delay and energy consumption over task offloading, UAV trajectory, IRS phase shifts, and computation resource allocation, and observes that the problem is a non-convex mixed-integer NP-hard program. To solve it online, the authors recast it as a Stackelberg game with MEC servers as leader and vehicles as followers. The follower level uses a many-to-one matching mechanism to produce feasible offloading decisions; the leader level uses a generative-diffusion-enhanced deep reinforcement learning algorithm for trajectory and IRS control, plus a KKT-based closed-form allocation of computation resources. Simulation results report a 2.5% lower average task completion delay and a 3.1% lower average energy consumption against the best benchmarks, with faster and more stable training than standard DRL baselines.","feed_headline":"Hierarchical optimizer cuts vehicular edge delay 2.5%, energy 3.1%","feed_subtitle":"A Stackelberg decomposition with diffusion-assisted reinforcement learning coordinates UAVs, IRSs, and servers online.","key_machinery":"The load-bearing mechanism is the Stackelberg-game decomposition of the original centralized MOOP into a leader problem (MEC servers) and a follower problem (vehicles). At the follower level, a many-to-one matching mechanism turns the discrete offloading decisions into stable, capacity-respecting associations between tasks and servers. At the leader level, the GDMTD3 algorithm—a deep reinforcement learning policy whose actor is a generative diffusion model that iteratively denoises continuous actions—produces UAV trajectory and IRS phase-shift decisions; a KKT-based closed-form solution resolves the computation-resource allocation exactly, removing those variables from the learned action spa","core_discovery":"The central claim is that the joint optimization of task offloading, UAV trajectory, IRS phase-shift configuration, and computation resource allocation—a mixed-integer non-linear program—can be solved online by exploiting the natural leader-follower hierarchy between the MEC servers and vehicles. The hierarchy is formalized as a Stackelberg game, and the authors show that the discrete offloading decisions can be generated by a matching mechanism, while the continuous trajectory and IRS decisions can be learned by a diffusion-enhanced actor-critic policy, with computation allocation computed in closed form from KKT conditions. The paper asserts this decomposition produces significant improvem","pith_inferences":["If the 2.5%/3.1% margins are real, the same leader-follower decomposition could transfer to other IRS-aided MEC settings—for example, multiple UAVs or aerial-only IRS fleets—with the matching and KKT modules reused unchanged.","The reported margins are small enough that unmodeled IRS phase-estimation overhead or channel aging in real deployments could absorb them; adding such overhead to the model is the most direct test of practical viability.","The paper does not disclose the weighting factors used in the simulation, so the exact delay-energy trade-off point is not reproducible from the text alone; a reader replicating the results would need those values.","The scalability experiments suggest the system's bottleneck shifts from computation to wireless contention as vehicle count grows, so coupling the matching mechanism with bandwidth allocation may delay the observed saturation of energy consumption."],"forward_implications":["If the reported gains hold, joint coordination of an aerial and a terrestrial MEC server with hybrid IRSs becomes a practical online control strategy for latency-sensitive vehicular applications in blockage-heavy urban settings.","Because computation allocation is closed-form, the learning agent only needs to output trajectory and phase decisions, which cuts the action dimension and memory footprint for online execution.","The Stackelberg reformulation implies that similar mixed discrete-continuous MEC problems can be approached hierarchically rather than as a monolithic DRL problem, potentially improving sample efficiency.","The robustness and scalability results indicate that the approach keeps working as the number of vehicles rises from 5 to 25 and as task sizes vary from 1 to 5 Mb.","The convergence comparison suggests that diffusion-based action generation stabilizes training in strongly coupled continuous action spaces compared with noise-based or entropy-based exploration baselines."],"fun_headline_variants":["Stackelberg game trims vehicular edge delay 2.5%, energy 3.1%","Diffusion-enhanced DRL cuts edge delay and energy in vehicular networks","IRS-aided MEC: hierarchical optimizer beats benchmarks on delay and energy","Online optimization for UAV-IRS MEC slashes delay and energy","Leader-follower MEC strategy lowers delay 2.5%, energy 3.1%"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The conclusion stands or falls on whether the simulated radio channels and the IRS's ability to steer signals accurately and without overhead match what happens in real cities.","fun_headline_variants_meta":{"raw":{"variants":["Stackelberg game trims vehicular edge delay 2.5%, energy 3.1%","Diffusion-enhanced DRL cuts edge delay and energy in vehicular networks","IRS-aided MEC: hierarchical optimizer beats benchmarks on delay and energy","Online optimization for UAV-IRS MEC slashes delay and energy","Leader-follower MEC strategy lowers delay 2.5%, energy 3.1%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000565,"raw_usage":{"total_tokens":2565,"prompt_tokens":840,"completion_tokens":1725,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":1618}},"tokens_in":584,"tokens_out":1725,"duration_ms":10800,"temperature":1.0,"reasoning_tokens":1618,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T13:16:58.629323+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a field test or a channel-accurate simulator, run the learned IRS phase-shift policy and then randomly permute the phase shifts while keeping the UAV trajectory and offloading decisions identical; if delay and energy do not measurably worsen, the IRS contribution is overstated. Alternatively, add a realistic phase-configuration overhead to the delay and energy objective—if that overhead exceeds the reported 2.5% and 3.1% margins, the claimed improvement disappears.","supporting_citations":[],"review_version":1}