{"id":"a1ffec72-bba5-473b-9dcf-1f07c423c8b7","arxiv_id":"2505.13331","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A centralized phasic policy gradient agent for elastic VR task offloading with multi-connectivity beats a decentralized independent agent by 28% in latency and 78% in energy in trace-driven simulations.","lead":"This paper designs three learning-based controllers that decide whether VR tasks run on the headset or are offloaded to an edge server over 4G, 5G, or WiGig links. In trace-driven simulations, the centralized controller used 78% less energy and 28% less latency than the decentralized one.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The shared-MEC constraint (11d) that motivates CPPG is never implemented in the simulator, so the reported 28%/78% advantage over IPPG may be a state-information artifact rather than evidence for centralized coordination.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing gap: the shared-MEC coupling that motivates centralized decision-making is not shown to be present in the evaluated environment. I agree with this assessment. The paper otherwise contains a plausible trace-driven study: real 4G/5G/WiGig throughput traces, a 360-degree video task-size distribution, a clearly formulated optimization objective, and three families of learning agents. The reported trends, such as CPPG avoiding 4G due to high energy and low throughput, are internally consistent with the stated cost model. However, the central claim is specifically that centralized full-observability decision-making dominates decentralized methods in a multi-user shared-edge setting. If the simulator does not instantiate the coupling in Eq. (11d), the optimization problem separates across users, and the claimed 28% latency and 78% energy reductions cannot be attributed to coordination over shared MEC resources. The paper itself credits the gap to full observability in Section VI, but full observability is only valuable when one user's state or action affects another user's reward. Without an explicit MEC allocation mechanism, queueing model, or capacity-sharing rule, the evaluation is incomplete at exactly the point where the conceptual contribution rests. This is not an internal inconsistency in the equations; it is an evaluation gap that a revision can address by specifying and implementing f^k_mec and re-running the comparison. No external baseline is included, but that is a secondary weakness; the more fundamental issue is whether the controlled comparison actually tests the claimed mechanism. The verdict should remain CONDITIONAL, conditional on the simulator instantiating Eq. (11d) and on the reported margins surviving that implementation.","tokens_in":20529,"tokens_out":2815,"duration_ms":32415,"concrete_test":"Run a controlled ablation in the simulator: (a) implement the shared-MEC constraint explicitly by defining an allocation rule for f^k_mec (e.g., proportional to offloaded workload or a single MEC queue with service rate Zmec) and retrain CPPG and IPPG with identical hyperparameters; (b) simultaneously run a degenerate environment where each user has a private, uncontended MEC frequency and Eq. (11d) is removed. If CPPG still beats IPPG by roughly 28%/78% in the private-MEC case, the reported advantage is an artifact of state information or implementation rather than shared-resource coordination. If the gap appears only when Eq. (11d) is active, the central claim is supported. Releasing the simulator code or a minimal reproducible script for this ablation would settle the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central justification for centralized decision-making is the multi-user coupling created by the MEC's limited computational resources. Eq. (11d) enforces sum_k z(f^k_mec) < Zmec, and Eq. (2)/Eq. (11b) define each user's offload response time through an allocated MEC frequency f^k_mec. However, neither the optimization section nor the evaluation section specifies how f^k_mec is computed, how Zmec is divided among concurrently offloading users, or whether any queueing or contention is modeled at the MEC. Section VI describes only trace-driven transmission rates and task sizes; it never states an allocation rule or service discipline. If the implemented environment instead gives each user an independent, uncontended MEC resource, then the per-user rewards are separable and the optimal action for user k depends only on that user's own state. In that case CPPG's full observability provides no principled advantage over IPPG, and the large reported gap likely reflects differences in architecture, state representation, or training rather than successful coordination under shared resources. Section V-A claims CPPG 'capture[s] the effect of shared computational resources on the reward each user receives,' but the paper supplies no simulation-level evidence that this coupling actually exists. Because the code and simulator are not released, this omission cannot be checked from the manuscript alone, and it directly undermines the interpretation of the headline 28% latency / 78% energy result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies elastic task offloading in a multi-user, multi-connectivity edge-assisted VR system. It formulates a constrained stochastic optimization problem that maximizes the sum of computational energy-efficiency across users, subject to per-user deadline constraints and a shared MEC computation capacity constraint (Eq. 11d). To solve this online, the authors propose three learning-based methods: a centralized phasic policy gradient (CPPG) agent with full observability, a decentralized independent phasic policy gradient (IPPG) agent, and decentralized shared multi-armed bandit variants (DSMAB). The agents are trained in a custom gym-like simulator driven by real 4G/5G/WiGig throughput traces and a 360-degree video dataset. The main reported result is that CPPG reduces latency by 28% and energy consumption by 78% compared to IPPG. The paper also analyzes the effect of task size, scalability with the number of users, and sensitivity to headset and MEC computational speeds.","tokens_in":20822,"tokens_out":6957,"duration_ms":63668,"significance":"If the reported gains are real and stem from the mechanism the authors invoke (centralized coordination under a shared MEC resource), the paper would provide a useful benchmark for centralized versus decentralized decision-making in edge-assisted immersive computing. The strengths are the use of real network traces and a public 360-degree video dataset, the formulation of an elastic-task model with a deadline-aware reward, and the inclusion of several decentralized alternatives for comparison. However, the central interpretation depends on the simulator faithfully instantiating the shared MEC capacity constraint, which is not demonstrated in the manuscript. As written, the headline comparison may reflect differences in state information and architecture rather than coordination benefits from handling shared computational resources, so the significance of the result is conditional.","major_comments":[{"comment":"The shared MEC capacity constraint that motivates centralized decision-making is not shown to be implemented in the simulator. Eq. (2) and Eq. (11b) define the per-user MEC frequency f^mec_k, and Eq. (11d) imposes sum_k z(f^mec_k) < Zmec, but Section VI, which describes the simulator, never specifies how Zmec is divided among concurrently offloading users, what value f^mec_k takes in the experiments, or whether any queueing or service discipline is modeled at the MEC. Since the reward for user k in Eq. (17) depends on f^mec_k only through that user's response time T_r^k, if f^mec_k is a fixed constant independent of the other users' actions, then the per-user rewards are separable and the environment contains no multi-user coupling. In that case the advantage of CPPG over IPPG would not be evidence of successful coordination under shared resources but rather a consequence of giving the centralized agent a richer state representation and a joint action space; the claim in Section V-A that CPPG 'capture[s] the effect of shared computational resources on the reward each user receives' would then be unsupported by the experiments. The authors should describe the MEC resource allocation rule (e.g., equal splitting, proportional demand, or a contention model), confirm that Eq. (11d) is actually enforced in the simulator, and re-run the comparison in an environment with genuine coupling before attributing the 28%/78% margins to centralized coordination.","section":"Sections III-A, IV, and VI (Eqs. (2), (11b), (11d))"},{"comment":"The deadline coefficient update contains a sign inconsistency. Eq. (14) defines L_{\\lambda_k} = \\lambda_k(T_d^k - T_r^k), so the gradient used in Eq. (25) is \\nabla_{\\lambda_k} L = T_d^k - T_r^k. When a task misses its deadline (T_r > T_d), this gradient is negative, and the update \\lambda_k \\leftarrow \\lambda_k + \\eta \\nabla_{\\lambda_k} L decreases \\lambda_k. The text states that \\lambda_k should increase when the response time exceeds the deadline, which is the opposite of what the formula implements. Standard dual ascent would use the constraint violation T_r - T_d with a positive step size. Please correct the sign (or the definition of L_{\\lambda_k}) and justify the resulting update. This is load-bearing because the reward in Eq. (17) depends on \\lambda_k, and the reported deadline violation probabilities in Section VI, Fig. 9a, are outputs of this procedure.","section":"Section V-A, Eqs. (14) and (25), Algorithm 1"},{"comment":"The passage from the constrained problem (11) to the Lagrangian form (12) and then to Eqs. (13)-(14) is not a valid duality argument as written. Minimizing \\lambda_k(T_d^k - T_r^k(u_k^*)) over \\lambda_k in Eq. (14) is degenerate: when the constraint is feasible the minimum is attained at \\lambda_k = 0, and when it is infeasible the expression is unbounded below for \\lambda_k \\ge 0. The paper should either provide a correct derivation of the Lagrangian dual (e.g., with \\lambda_k \\ge 0 and the standard saddle-point formulation) or explicitly state that the deadline penalties are a heuristic mechanism rather than the exact dual of Eq. (11). Since the training algorithm and the reported constraint-satisfaction results are built on this formulation, the distinction is consequential for the correctness of the method.","section":"Section IV, Eqs. (11)-(14)"}],"minor_comments":[{"comment":"There is a typo in 'elsatic tasks' in the first paragraph of the Introduction; it should read 'elastic tasks'.","section":"Section I"},{"comment":"'refereed to' appears twice (for ILin-UCB and ILin-TS) and should be 'referred to'.","section":"Section V-C"},{"comment":"'L Jonit' is a typo for 'L Joint' in both equations.","section":"Section V-A, Eqs. (23) and (29)"},{"comment":"The DRL training hyperparameters (learning rates, batch sizes, optimizer, clip epsilon, discount factor) are not listed; without them the experiments are not reproducible. Please add a complete hyperparameter table.","section":"Section VI, Table I"},{"comment":"The state size for IPPG is stated as 15, but the composition of the 15 features (e.g., the history length used for transmission time, response time, and energy consumption) is not specified. Please define the state vector explicitly.","section":"Section VI, second paragraph"},{"comment":"The markers for the different methods are difficult to distinguish because several points cluster in the same region. Consider reporting a table of mean and standard deviation values, or labeling the points directly.","section":"Section VI, Figure 8"},{"comment":"The 28% latency and 78% energy reductions are stated without the experimental conditions (e.g., K=30, specific traces, and testing episodes). Please add a qualifier or refer the reader to the exact setup in Section VI.","section":"Abstract and Section VII"}],"recommendation":"major_revision","confidential_remarks":"The central issue to resolve in revision is the missing implementation of the shared MEC computation constraint. The paper's stated contribution is the centralized-versus-decentralized comparison under shared resources, but without an implemented allocation rule the reported advantage of CPPG may be an artifact of the state representation rather than a genuine coordination gain. The deadline-coefficient sign error also needs fixing, as it directly affects the training procedure. If the authors can show that the environment couples users through Eq. (11d) and that the CPPG advantage persists under that coupling, the paper would be significantly stronger."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid architecture study, but the headline result is contingent on a simulation detail that is never specified. The stress-test note is right. Eq. (11d) imposes a sum constraint on MEC compute, but Section VI never says how f^mec_k is allocated among users or whether the simulator actually enforces any contention. If each user gets an independent, uncontended MEC slice, then the multi-user coupling that motivates CPPG does not exist, and the 28%/78% gap is just a centralized-oracle vs partial-observation artifact.\n\nWhat is genuinely new: the combination of elastic quality-layer VR tasks, multi-connectivity offloading, and a centralized phasic policy gradient solver, plus independent PPG and contextual-bandit variants, is not in the prior work I know. The task-size policy analysis (more offloading as tasks grow, avoidance of the 4G channel) is a clean empirical result. The use of real 4G/5G/WiGig traces and a 360-degree video dataset is a plus. The optimization formulation is coherent and the reward model is explicit.\n\nSoft spots, in proportion. The missing MEC allocation is the big one. It is not a mathematical error; it is an unspecified implementation detail that directly bears on the interpretation of the central result. The paper claims CPPG \"captures the effect of shared computational resources,\" but no simulation-level evidence shows that coupling actually exists. Second, there are no external or prior-art baselines; every comparison is self-generated in the authors' simulator. Third, the headline margins have no confidence intervals or significance tests. Fourth, the code and simulator are not released, so none of this can be checked from the manuscript alone.\n\nWhere that leaves us: the paper is worth reading as a design study, and the task-size results may well be robust. But the 28% latency / 78% energy claim should be treated as provisional until the MEC allocation mechanism is specified or the simulator is released.\n\nMy recommendation: send it to peer review, but require the authors to state exactly how f^mec_k is computed and whether Eq. (11d) is enforced in the environment. If the coupling is absent, they need to soften the interpretation and rerun with a true shared-resource model. This is fixable, and the rest of the paper is competent enough to justify the effort.","headline":"A competent centralized-offloading DRL paper whose headline 28% latency / 78% energy gains likely rest on an unimplemented shared-MEC constraint; treat the margins as provisional until the allocation mechanism is specified.","tokens_in":21364,"tokens_out":2537,"would_cite":false,"duration_ms":23479,"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 centralized phasic policy-gradient agent that sees all users jointly cuts VR offloading latency by 28% and energy consumption by 78% compared with independent per-user agents.","keywords":["elastic task offloading","multi-connectivity VR systems","edge computing","computational energy-efficiency","multi-agent reinforcement learning","centralized versus decentralized decision-making","phasic policy gradient","360-degree video"],"falsifier":"Instrument the offloading simulator to record, for every task, the MEC frequency $f^k_{\\mathrm{mec}}$ actually assigned to each user, and rerun the comparison with the shared constraint $\\sum_k z(f^k_{\\mathrm{mec}}) < Z_{\\mathrm{mec}}$ replaced by independent per-user MEC caps. If CPPG still shows roughly 28% lower latency and 78% lower energy when no user's offloading affects another's compute allocation, the result is an observability artifact; if the gap disappears, shared-resource coordination is doing the work.","tokens_in":20281,"feed_emoji":"🥽","tokens_out":9132,"duration_ms":83744,"temperature":0.7,"pith_summary":"The paper argues that when VR users with elastic computational tasks share an edge server and several wireless channels, offloading decisions are best made by a single centralized agent that sees every user's task and history, rather than by independent per-user agents. The authors formulate the problem as a constrained stochastic optimization that maximizes computational throughput per energy unit while respecting task deadlines, and solve it online with a centralized phasic policy gradient (CPPG) agent. On a trace-driven simulator with 30 users, real 4G/5G/WiGig throughput traces, and a 360-degree video rendering task set, they report that CPPG reduces response latency by 28% and energy consumption by 78% relative to the independent phasic policy gradient (IPPG) baseline, at the price of full observability. If the comparison holds, it means that coordination through shared edge resources, not just per-user channel selection, is where the largest gains in immersive edge computing lie.","feed_headline":"Central agent cuts VR task latency 28% and energy 78%","feed_subtitle":"One edge-side agent decides offloading for every VR user, trading full observability for far lower latency and energy.","key_machinery":"The load-bearing mechanism is the constrained computational energy-efficiency objective of Eq. (11), paired with the CPPG actor-critic architecture. Phasic policy gradient here means a PPO-style training loop with a separate auxiliary phase: after the usual policy and value update, the network is further trained on behavioral-cloning and value losses, which the paper argues reduces interference and lets the joint state shape each user's reward. The centralized actor takes a joint action $[u^1, \\dots, u^K]$, while the critic emits per-user state values. The per-user reward folds energy efficiency and deadline compliance into one scalar, so each environment step is self-contained and no bootstrapping is needed. The shared-MEC constraint $\\sum_k z(f^k_{\\mathrm{mec}}) < Z_{\\mathrm{mec}}$ is what makes the joint decision matter in principle: one user's offload changes the edge compute available to others, and CPPG is designed to capture that coupling.","core_discovery":"The paper's central claim is that a centralized multi-agent reinforcement-learning agent, CPPG, can solve the constrained stochastic computational energy-efficiency maximization problem for multi-connectivity VR task offloading in real time. CPPG observes the full system state, namely each user's elastic task size and computational intensity plus past transmission times, response times, and energy consumption, and outputs a joint offloading action for all users, choosing local execution or one of the available channels (4G, 5G, or WiGig) to a shared MEC unit. The reward for each user is the task's computational throughput per joule plus a learned deadline-violation penalty, and because this reward is self-contained the discount factor can be set to zero. After training on real network traces and 360-degree video data, the authors report that CPPG outperforms the decentralized IPPG baseline by 28% in latency and 78% in energy consumption, and that the learned policy offloads larger tasks, favors 5G and WiGig over 4G, and shifts computation toward whichever side, headset or MEC, has more available speed.","pith_inferences":["If the claimed 28%/78% edge comes mostly from full observability, a decentralized agent given a compact aggregate signal, such as the MEC's current load, could recover part of that gap; the paper does not test this.","Because the reward is a per-task scalar with no temporal dependency, the centralized problem is closer to contextual bandits than to full RL; comparing CPPG against a contextual bandit with the same joint state would isolate what the phasic policy-gradient machinery adds.","The shared-MEC constraint of Eq. (11d) is asserted, but the simulator's allocation of MEC frequency among users is not specified; if users do not actually contend for one MEC, the reported advantage is an information advantage by construction.","Task elasticity could be turned into an active control: instead of taking task size as given, an agent could select the video-quality layer jointly with the offloading action, extending energy efficiency to a rate-quality trade-off frontier."],"forward_implications":["A single edge-side controller can serve many VR users with one inference pass: CPPG's decision time is $O(1)$ in the number of users, while decentralized IPPG scales as $O(K)$.","Per-user, partially observable offloading is materially worse on this workload: the reported numbers imply 28% higher latency and 78% higher energy consumption compared with centralized joint decisions.","The learned policy is task-size aware: larger elastic tasks are increasingly offloaded rather than computed locally, and higher-throughput channels (5G, WiGig) absorb the offloaded load while 4G is avoided.","System resource changes steer the policy in predictable directions: faster VR headsets pull computation local, faster MEC pulls it to the edge, and both moves improve response time.","Quality-of-experience targets are reachable: in the 30-user test setting the paper reports 12.24% of tasks missing the one-second deadline overall, or 0.4% per user."],"supporting_citations":[{"why":"Supplies the phasic policy gradient method that both CPPG and IPPG build on.","marker":"[6]"},{"why":"Supplies the CPU dynamic-power model used to compute local computation energy in Eq. (6).","marker":"[50]"},{"why":"Supplies the linear frequency-voltage relationship that converts CPU frequency into energy in Eq. (6).","marker":"[51]"},{"why":"Provides the real 4G/5G throughput traces and power allocation profile used in training and evaluation.","marker":"[54]"},{"why":"Supplies the dual-clip PPO loss used to update the actor and critic networks.","marker":"[57]"},{"why":"Supports IPPG's parameter-sharing training, making the decentralized baseline comparable in sample efficiency.","marker":"[59]"},{"why":"Provides the 360-degree video segments and layer-based task sizes that define the elastic task distribution.","marker":"[61]"}],"fun_headline_variants":["VR offloading: one central agent cuts latency 28%, energy 78%","Centralized RL for VR tasks: 28% less latency, 78% less energy","Central agent for elastic VR offloading: 78% energy saved, 28% latency","Multi-connectivity VR: central RL offloading beats decentralized by 28%/78%","Elastic VR tasks: central agent learns to cut energy 78%, latency 28%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline comparison assumes that the users genuinely share one constrained edge-computing resource during the experiments; if each user effectively gets independent edge service, the reported CPPG advantage would come from central observability rather than from coordination.","fun_headline_variants_meta":{"raw":{"variants":["VR offloading: one central agent cuts latency 28%, energy 78%","Centralized RL for VR tasks: 28% less latency, 78% less energy","Central agent for elastic VR offloading: 78% energy saved, 28% latency","Multi-connectivity VR: central RL offloading beats decentralized by 28%/78%","Elastic VR tasks: central agent learns to cut energy 78%, latency 28%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000974,"raw_usage":{"total_tokens":4213,"prompt_tokens":1093,"completion_tokens":3120,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":709,"completion_tokens_details":{"reasoning_tokens":3006}},"tokens_in":709,"tokens_out":3120,"duration_ms":21879,"temperature":1.0,"reasoning_tokens":3006,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:16:29.314622+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Instrument the offloading simulator to record, for every task, the MEC frequency $f^k_{\\mathrm{mec}}$ actually assigned to each user, and rerun the comparison with the shared constraint $\\sum_k z(f^k_{\\mathrm{mec}}) < Z_{\\mathrm{mec}}$ replaced by independent per-user MEC caps. If CPPG still shows roughly 28% lower latency and 78% lower energy when no user's offloading affects another's compute allocation, the result is an observability artifact; if the gap disappears, shared-resource coordination is doing the work.","supporting_citations":[{"cited_title":"Energy-Optimal Mobile Cloud Computing under Stochastic Wireless Channel,","cited_arxiv_id":null,"evidence_quote":"Supplies the CPU dynamic-power model used to compute local computation energy in Eq. (6)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the linear frequency-voltage relationship that converts CPU frequency into energy in Eq. (6)."},{"cited_title":"A Variegated Look at 5G in the Wild: Performance, Power, and QoE Implications,","cited_arxiv_id":null,"evidence_quote":"Provides the real 4G/5G throughput traces and power allocation profile used in training and evaluation."},{"cited_title":"Full UHD 360-Degree Video Dataset and Modeling of Rate- Distortion Characteristics and Head Movement Navigation,","cited_arxiv_id":null,"evidence_quote":"Provides the 360-degree video segments and layer-based task sizes that define the elastic task distribution."}],"review_version":1}