{"id":"2600214c-ddb8-4c10-9d9e-e07eb3c7df45","arxiv_id":"2502.08119","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A GAI-enhanced HAPPO algorithm for joint task offloading and UAV trajectory planning in a UAV-ground station MEC network serving USVs, claiming a 22.8% delay improvement over MARL baselines.","lead":"This paper proposes a generative-AI-enhanced multi-agent reinforcement learning algorithm to coordinate drones and ground stations helping unmanned surface vehicles offload computing tasks. The authors report a 22.8% improvement over existing MARL baselines in simulations, though the underlying system model contains errors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The UAV/GS queue updates in Eqs. (20)–(21) corrupt the simulated environment, so the 22.8% performance claim is not evidence for GAI-HAPPO on the stated problem.","rationale":"The paper proposes GAI-HAPPO, a transformer/GAN-enhanced MARL algorithm for joint task offloading and UAV trajectory planning, and claims a 22.8% performance improvement over benchmarks. The only evidence for this claim is the simulation study in Section IV. The simulation environment is specified by the queue dynamics in Section II.D, the delay equations in Eqs. (22)–(25), and the reward in Eq. (33). Eq. (20) multiplies the existing UAV queue backlog by sum_i p_{i,j}(t), the number of USVs offloading to that UAV, which resets the backlog whenever no USV offloads and doubles or triples it when several do. It also adds beta_i(t)d_i(t) to every UAV queue for every USV, independent of the offloading decision. Eq. (21) does the same for ground stations. Because these queue values propagate into the delay and reward, the MARL agents are trained against a corrupted transition and reward model. This is not a cosmetic typo: it changes the optimization landscape and can arbitrarily inflate or deflate the measured reward, so the reported ranking of algorithms does not establish that GAI-HAPPO solves problem P0. The reader identified the same equation and the same consequence, so I agree with the rejection. The only caveat is that this concern could in principle be resolved by rerunning the simulations with corrected recursions; however, as the paper currently stands, the central claim is unsupported. The absence of code, seed counts, and error bars compounds the problem but is secondary to the model error.","tokens_in":8788,"tokens_out":5489,"duration_ms":101668,"concrete_test":"Implement two versions of the simulation: (1) the queue recursions exactly as printed in Eqs. (19)–(21), and (2) corrected recursions of the form Q^u_j(t+1) = max{0, Q^u_j(t) + sum_i p_{i,j}(t)beta_i(t)d_i(t) - tau f^u_j} and Q^g_k(t+1) = max{0, Q^g_k(t) + sum_i q_{i,k}(t)gamma_i(t)d_i(t) - tau f^g_k}, with all other equations, parameters, and hyperparameters unchanged. Rerun the Section IV comparisons with at least five seeds and check whether GAI-HAPPO still produces the Fig. 3 training curves and the reported 22.8% improvement. As a minimal analytical check, take one USV, one UAV, and a time slot with p_{1,1}(t)=0 and beta_1(t)>0: the printed Eq. (20) resets the UAV backlog to zero, while the corrected recursion retains the existing backlog and adds beta_1(t)d_1(t); the next-slot delay and reward differ, proving the printed environment is not the stated one.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that GAI-HAPPO outperforms benchmarks by 22.8% rests entirely on simulations whose environment is defined by the queue recursions, delay equations, and reward. Eq. (20) defines Q^u_j(t+1) = max{0, sum_i p_{i,j}(t)Q^u_j(t) + beta_i(t)d_i(t) - tau f^u_j}. This is not a queue update: the existing backlog Q^u_j(t) is multiplied by the number of USVs offloading to UAV j, so the backlog is deleted when no USV offloads to that UAV and scaled up when several do. Moreover, beta_i(t)d_i(t) is added to every UAV queue for every USV i, regardless of whether USV i offloads to that UAV or to any UAV at all, because beta_i(t) is not multiplied by p_{i,j}(t). Eq. (21) has the same structure for GS queues with gamma_i(t)d_i(t) and q_{i,k}(t). These corrupted queues enter the UAV and GS computation delays in Eqs. (23)–(24), the total delay in Eq. (25), and the reward r(t) = sum_i d_i(t)/T^t_i(t) in Eq. (33). Consequently, every state transition and reward signal used to train GAI-HAPPO and the benchmark algorithms is generated by a system model inconsistent with the stated problem P0. The reported convergence in Fig. 3 and the 22.8% improvement are therefore not a valid test of the proposed algorithm for the formulated MEC problem. This load-bearing algebraic error is independent of the additional lack of code, seeds, and error bars.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers a maritime multi-access edge computing (MEC) scenario in which unmanned surface vehicles (USVs) generate computation tasks that can be processed locally, by UAVs, or by ground stations. It formulates a mixed-integer nonlinear program (MINLP) that jointly optimizes task offloading, workload splitting, and UAV trajectories to minimize total execution time, and then proposes GAI-HAPPO, a heterogeneous-agent proximal policy optimization algorithm that uses a transformer-based actor and a GAN-based critic. Simulations compare GAI-HAPPO with HAA2C, HAPPO, GAN-HAPPO, and Transformer-HAPPO, reporting a 22.8% improvement in average reward. The core of the paper is the system model in Section II and the algorithm design in Section III, followed by simulation results in Section IV.","tokens_in":9161,"tokens_out":3944,"duration_ms":32621,"significance":"If the proposed algorithm and simulation results were valid, the work would be a useful contribution to latency-aware task offloading for USV networks, and the combination of transformer actors and GAN critics within a HAPPO framework is a timely algorithmic idea. The paper also benefits from a reasonably broad benchmark comparison. However, the central queue dynamics in the system model are algebraically inconsistent, and this error propagates through the delay and reward functions that define the simulation environment. The reported performance gains therefore do not currently provide valid evidence for the proposed approach on the stated problem. The contribution is not established as presented, and the core issue would require a re-derivation of the model and a full re-running of the experiments.","major_comments":[{"comment":"The UAV queue recursion is algebraically inconsistent with the intended backlog model. The term \\sum_i p_{i,j}(t) Q_j^u(t) multiplies the existing queue by the offloading indicator sum, so the backlog is erased whenever no USV offloads to UAV j in slot t, and it is scaled by the number of offloading USVs when several do. In addition, \\beta_i(t)d_i(t) is summed over all i for every UAV j regardless of whether USV i offloads to j, so work that is not sent to that UAV still enters its queue. Equation (21) has the same defect for the ground-station queue. Because these queues enter the computation delays in Eqs. (23)-(24), the total delay in Eq. (25), and the reward in Eq. (33), every state transition and reward signal used in training and evaluation is generated by an environment that does not correspond to problem P0. The reported 22.8% improvement and the convergence behavior in Fig. 3 are therefore not evidence for GAI-HAPPO on the stated problem.","section":"Section II-D, Eq. (20)"},{"comment":"The USV action space is defined as a vector of real-valued variables containing p_{i,j}, q_{i,k}, \\alpha_i, \\beta_i, and \\gamma_i, but p and q are binary by Eqs. (14)-(15) and must satisfy the exclusive-offload constraints (16)-(17). The paper does not describe how continuous actions are mapped to feasible binary decisions, nor whether the constraints are enforced by masking, projection, or penalty. Without such a mechanism, the learned policy cannot be guaranteed to produce admissible offloading choices, and the comparison with benchmark algorithms is not well-defined.","section":"Section III-A, Eq. (31)"},{"comment":"The simulation results reuse the corrupted queue environment for both GAI-HAPPO and all baselines, so the relative ordering of the algorithms may be an artifact of the erroneous dynamics. The paper claims a 22.8% improvement in the abstract, introduction, and conclusion, but Section IV never defines how this number is computed, which baseline it refers to, or which figure supports it. Furthermore, the training curves in Fig. 3 and the bar charts in Fig. 4 are presented without multiple seeds or error bars, making it impossible to assess statistical significance even if the environment equations were corrected.","section":"Section IV-B and IV-C"}],"minor_comments":[{"comment":"The path-loss formula contains a typo: the term \"exp{−b[ςi,j(t) − a}\" is missing a closing bracket, and the denominator \"1 + a exp{...}\" is ambiguous. Please correct the expression and ensure all brackets are balanced.","section":"Section II-A, Eq. (1)"},{"comment":"In the description after Eq. (31), the text says \"the UAV needs to choose which UAV and GS to offload the tasks,\" but this action space is for the USV agent. This appears to be a typo and should be fixed.","section":"Section III-A, action space"},{"comment":"The hyperparameter settings are incomplete: only learning rates are given (5×10^{-5} and 10^{-4} for actor and critic, presumably), but the PPO clip range, batch size, discount factor, and GAN training details are not specified. Please provide a complete table of parameters to support reproducibility.","section":"Section IV-A"},{"comment":"The introduction and related discussion rely heavily on the authors' prior works [2], [4], [6], and [14]. The novelty of the proposed framework relative to these prior contributions should be stated more explicitly in the introduction.","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper is built around a load-bearing algebraic error in the queue model that invalidates the simulation evidence. The issue is not merely a presentation defect: it changes the state transitions, the delays, and the rewards, and it affects every experimental result. Even if the authors fixed Eqs. (20)-(21), they would need to rerun all simulations and re-evaluate the 22.8% claim. The lack of code, seeds, and error bars further weakens the empirical support. The topic is relevant to the journal's scope, but the manuscript in its current form does not meet the standard for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the 22.8% improvement claim is not supported because the queue recursions in Eqs. (20)–(21) corrupt the simulated environment. That’s the main thing you should know.\n\nThe paper does have a real contribution at the architectural level: combining a transformer-based actor with a GAN-based critic inside the HAPPO sequential update scheme, applied to UAV/ground-station MEC for USVs. That specific combination is new as far as I can tell, and the ablation baselines (GAN-HAPPO, Transformer-HAPPO) are the right way to isolate the two components. The system model and problem formulation are standard MINLP fare, and the writing is clear enough.\n\nThe soft spot is not small. Eq. (20) defines Q^u_j(t+1) = max{0, sum_i p_{i,j}(t) Q^u_j(t) + beta_i(t)d_i(t) - tau f^u_j}. The existing backlog is multiplied by the number of offloading USVs, so the queue resets to zero whenever no USV offloads to that UAV. And beta_i(t)d_i(t) is added to every UAV queue regardless of whether USV i offloads there (or anywhere). Eq. (21) has the same structure for GSs. The queues feed directly into the delays in Eqs. (23)–(24), the total delay, and the reward in Eq. (33). So every transition and reward used to train GAI-HAPPO and the benchmarks is generated by a model that is not the one described in problem P0. That is load-bearing, not a cosmetic typo.\n\nThere is also no code, no seeds, and no error bars; the 22.8% figure appears in the abstract and conclusion but I couldn't find a statistical test or a direct point comparison in the figures. The binary offloading variables are placed in a continuous action space with no feasibility mechanism, which is a secondary concern.\n\nOn the citation pattern: they rely heavily on their own prior work, but those references are relevant and standard in this subfield, so I wouldn't call it a problem.\n\nWho gets value from this? Readers working on MARL architectures for MEC might look at the actor-critic design, but the empirical validation is not usable in its current form. I would not send this to peer review as is — a desk reject with clear instructions to fix the queue recursions and rerun all simulations is the appropriate path. If the authors correct the model and the results still show an improvement, then it becomes a reasonable incremental paper.","headline":"Load-bearing queue update error in Eq. (20) invalidates the simulation results, so the 22.8% claim is not evidence; the architecture combination is real but needs a corrected model.","tokens_in":9661,"tokens_out":3565,"would_cite":false,"duration_ms":27498,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A generative-AI-enhanced multi-agent reinforcement learning algorithm, GAI-HAPPO, is claimed to reduce total task execution time by 22.8% in a cooperative UAV-ground-station MEC system serving unmanned surface vehicles.","keywords":["unmanned surface vehicles","multi-access edge computing","task offloading","UAV trajectory planning","multi-agent reinforcement learning","heterogeneous agent proximal policy optimization","generative adversarial network","transformer"],"falsifier":"Recompute the simulated delays, rewards, and training curves after replacing the UAV queue update with one that preserves unfinished work, e.g., $Q^u_j(t+1)=\\max\\{0, Q^u_j(t) + \\sum_i p_{i,j}(t)\\beta_i(t)d_i(t) - \\tau f^u_j\\}$. If GAI-HAPPO no longer shows the reported 22.8% advantage over the same baselines, the performance claim rests on the current queue formulation rather than on the algorithm alone.","tokens_in":8589,"feed_emoji":"🚢","tokens_out":5447,"duration_ms":44239,"temperature":0.7,"pith_summary":"This paper tries to establish that a generative-AI-enhanced multi-agent reinforcement learning algorithm, GAI-HAPPO, can jointly schedule task offloading and UAV flight trajectories so that unmanned surface vehicles in a maritime edge-computing network finish computation faster. The authors formulate the joint problem as a mixed-integer nonlinear program, then solve it with transformer-based actor networks and a GAN-stabilized critic inside the HAPPO training scheme. If the claim is right, the algorithm gives a ready method for coordinating USVs, UAVs, and ground stations under task and trajectory uncertainty, with a reported 22.8% performance improvement over four benchmarks.","feed_headline":"AI-enhanced fleet coordination cuts maritime compute delays 22.8%","feed_subtitle":"Generative AI inside a multi-agent RL policy jointly plans UAV paths and task offloading, beating four benchmarks.","key_machinery":"The central mechanism is the GAI-HAPPO training loop: heterogeneous agents (USVs deciding offloading fractions and UAVs deciding flight azimuth and distance) are updated sequentially under a clipped surrogate objective, following HAPPO. A transformer self-attention layer in the actor maps state inputs to actions, and a GAN-based critic replaces the usual value network, with a generator predicting state values and a discriminator distinguishing true from predicted values. The advantage function is computed by generalized advantage estimation and propagated through the sequential updates.","core_discovery":"The paper's central claim is that GAI-HAPPO outperforms existing benchmark algorithms—HAA2C, HAPPO, GAN-HAPPO, and Transformer-HAPPO—in minimizing total task execution time for a cooperative UAV-and-ground-station MEC system serving USVs. The authors report that the proposed algorithm achieves the highest average reward with stable convergence after 2,500 episodes, consistently lower delays as the numbers of USVs and UAVs vary, and an overall 22.8% performance improvement. The improvement is attributed to the transformer's capacity to model complex state dependencies and the GAN-based critic's contribution to training stability.","pith_inferences":["A natural test that the paper does not run is a component ablation: replacing the transformer actor or the GAN critic with plain MLPs under identical seeds would isolate which generative component causes the 22.8% gain.","The queue update in Eq. (20) multiplies the existing UAV queue by the offloading indicators, so the backlog resets whenever no USV offloads to that UAV; a corrected backlog equation could change the simulated delays and rewards, though the algorithmic contribution would still be testable.","Because the reward is a throughput-like ratio (data size divided by delay), the method targets latency per bit; applying it to energy or fairness objectives would require a new reward design but the same actor-critic machinery."],"forward_implications":["If GAI-HAPPO works as reported, task offloading and UAV trajectory planning can be solved jointly in a single learned policy for USV-UAV-GS networks.","The algorithm's stable convergence after 2,500 episodes suggests it can be deployed in dynamic maritime environments where task arrivals and USV trajectories keep changing.","The reported 22.8% improvement over HAPPO means the generative components add practical value beyond the base heterogeneous-agent method.","The same training structure could be applied to other MEC settings where agents have heterogeneous action spaces, such as ground robots and aerial relays."],"supporting_citations":[{"why":"Supplies the HAPPO sequential update scheme and the HAA2C/HAPPO baselines used for comparison.","marker":"[16]"},{"why":"Supplies the generative-AI-for-DRL framework that motivates the transformer actor and GAN critic.","marker":"[12]"},{"why":"Provides the environment-dependent path loss parameters used in the USV-to-UAV channel model.","marker":"[14]"},{"why":"Supplies the Gauss-Markov mobility model used for USV trajectory dynamics.","marker":"[15]"}],"fun_headline_variants":["GenAI-enhanced UAV-GS MEC cuts USV task delays by 22.8%","GAI-HAPPO beats four baselines, slicing maritime compute time 22.8%","Generative AI plus multi-agent RL cut USV edge delays 22.8%","22.8% faster maritime compute via GenAI-coordinated UAVs and ground stations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the UAV queue update in Eq. (20) correctly describes the backlog of tasks waiting on each UAV; as written it resets the backlog to zero whenever no USV offloads to that UAV and scales it by the number of offloading USVs otherwise, so the delays and rewards used in training and evaluation depend on this equation being an accurate model.","fun_headline_variants_meta":{"raw":{"variants":["GenAI-enhanced UAV-GS MEC cuts USV task delays by 22.8%","GAI-HAPPO beats four baselines, slicing maritime compute time 22.8%","Generative AI plus multi-agent RL cut USV edge delays 22.8%","22.8% faster maritime compute via GenAI-coordinated UAVs and ground stations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000649,"raw_usage":{"total_tokens":2971,"prompt_tokens":929,"completion_tokens":2042,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":1947}},"tokens_in":545,"tokens_out":2042,"duration_ms":14767,"temperature":1.0,"reasoning_tokens":1947,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T10:21:59.140440+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the simulated delays, rewards, and training curves after replacing the UAV queue update with one that preserves unfinished work, e.g., $Q^u_j(t+1)=\\max\\{0, Q^u_j(t) + \\sum_i p_{i,j}(t)\\beta_i(t)d_i(t) - \\tau f^u_j\\}$. If GAI-HAPPO no longer shows the reported 22.8% advantage over the same baselines, the performance claim rests on the current queue formulation rather than on the algorithm alone.","supporting_citations":[{"cited_title":"Heterogeneous-agent reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the HAPPO sequential update scheme and the HAA2C/HAPPO baselines used for comparison."},{"cited_title":"Joint HAP access and LEO satellite backhaul in 6G: Matching game-based approaches,","cited_arxiv_id":null,"evidence_quote":"Provides the environment-dependent path loss parameters used in the USV-to-UAV channel model."},{"cited_title":"Mobility models, traces and impact of mobility on opportunistic routing algorithms: A survey,","cited_arxiv_id":null,"evidence_quote":"Supplies the Gauss-Markov mobility model used for USV trajectory dynamics."}],"review_version":1}