{"id":"9c2ffcc6-6250-459b-83ba-25879f6adc27","arxiv_id":"2502.02886","paper_version":3,"verdict":"UNVERDICTED","confidence":"MODERATE","novelty_score":0.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A reading report on MEC and Open RAN concludes that AI and deep reinforcement learning improve offloading and network control, but it provides no new measurements.","lead":"This paper is a reading report on two wireless networking ideas: putting computing power closer to phones, and opening up the radio network so different vendors can work together. It says artificial intelligence can make both systems run better, but it only repeats findings from earlier papers and adds no new tests.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quantified performance claims in Sections III.B and IV.B are attached to references that do not contain them, so the central 'AI/ML improves performance' assertion is not verifiable from the cited sources.","rationale":"This is a reading report and survey, not an original research paper, so there is no new technical claim to accept or reject. The appropriate verdict is UNVERDICTED, as the reader concluded. My stress-test focuses on the one component that would make the survey trustworthy: the accuracy and attribution of the numerical performance improvements that support the paper's central assertion that AI/ML materially improves MEC and ORAN. The reader's weakest_assumption identified exactly this issue ('the performance numbers quoted from cited papers are accurate, correctly attributed, and transferable'), and I agree with that identification, which is why I did not raise a different concern. I chose 'partial' rather than 'agree' because the reader also mentioned transferability to production networks, whereas my check is narrower: the numbers as stated are not even traceable to the cited references. The concrete test is straightforward and does not require re-running simulations: it only requires a careful comparison of each claimed percentage against the text of the cited paper. If the numbers check out, the survey's summary is accurate. If they do not, the report's central evidence collapses to qualitative statements about DRL being promising, which are far weaker than the quantified claims currently presented. Either way, the verdict remains UNVERDICTED because there is no original result to validate; the only possible change would be a more cautious presentation of the quantitative benefits. I see no internal inconsistency in the paper's argument, and I give credit for the paper being transparent about its nature as a reading report. The concern is not about novelty or style but about the reliability of the evidence chain, which is the basis for any claim that the report is a useful secondary source. The proposed test would settle whether the report is a trustworthy summary or an unreliable one, and it can be done with the cited literature in hand.","tokens_in":7566,"tokens_out":2676,"duration_ms":25368,"concrete_test":"For each claimed metric in Sections III.B and IV.B, locate the exact sentence in the cited paper (or corrected reference) that reports that number. Specifically verify: (a) the DROO paper (IEEE TMC 2020, ref [1]) reports a 23% energy reduction and a <5% task-failure rate versus the stated heuristic baseline; (b) the traffic-steering paper (arXiv:2209.14171, ref [10]) reports a 17% load-balancing improvement; (c) the ColO-RAN paper (IEEE TNSM 2021, ref [8]) reports a 21% task-failure reduction; and (d) the O-RAN Performance Analyzer (IEEE Comm Mag 2024, ref [9]) reports a 22% congestion reduction. If any number is absent, stated in a different scenario, or comes from a different reference, mark that claim as unverified and adjust the review's quantitative summary accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that AI/ML, especially DRL, materially improves MEC and ORAN systems, and the only quantitative evidence offered is a set of percentage improvements: DROO's 23% energy reduction, 17% task-completion-time reduction, and <5% task-failure rate in Section III.B, and the ORAN xApps' 17% load-balancing, 21% task-failure reduction, 25% interference reduction, 18% handover improvement, and 22% congestion reduction in Section IV.B. These numbers are load-bearing because the qualitative conclusion that AI/ML is beneficial rests on them. However, the citations do not line up with the claims. In Section III.B, the 17% comparison to an 'MDP-based energy-harvesting approach' cites references [1][3], but [3] is Chiang and Zhang, 'Fog and IoT,' which is not an MDP energy-harvesting study. The task-failure comparison to heuristics cites [1][2], yet [2] is 'Wireless Powered Communication: Opportunities and Challenges,' not a heuristic offloading baseline. In Section IV.B, the 17% load-balancing figure from the traffic-steering work is attributed to reference [7], which is the X5g testbed paper; the actual traffic-steering paper is reference [10]. The ColO-RAN 21% task-failure reduction is attributed to [9], the O-RAN Performance Analyzer, rather than to reference [8], which is the ColO-RAN paper. Similarly, the 22% congestion reduction is attributed to [8], while the Performance Analyzer is [9]. Thus, a reader cannot verify the central quantitative claims without first correcting the reference list, and the abstracted numbers may mix results from different experiments or papers. This does not make the review fraudulent, but it does mean the report's evidence for its main assertion is currently unsupported hearsay, and the UNVERDICTED verdict is the right disposition.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a literature survey (self-described as a 'reading report') reviewing recent work on AI/ML, particularly deep reinforcement learning (DRL), for computation offloading in mobile edge computing (MEC) and for closed-loop control via xApps in Open RAN (ORAN). It describes the DROO algorithm, reports quantitative gains over heuristic and MDP-based baselines, surveys ORAN architecture, and summarizes performance improvements from several xApps for traffic steering, network slicing, interference management, handover, and congestion avoidance. The paper contains no new experiments, derivations, or original data; its contribution is a synthesis of selected prior work and a discussion of limitations and future directions.","tokens_in":1481,"tokens_out":1536,"duration_ms":50029,"significance":"If accurate, this survey would serve as a compact entry point to DRL for MEC and ORAN, particularly for its explicit enumeration of DROO's limitations (feedback dependence, exploration-exploitation trade-offs, fault tolerance) and its organization around self-healing ORAN use cases. The paper's evidentiary value rests entirely on faithfully reporting cited results, and the citation mismatches detailed below currently prevent verification of the central quantitative claims. The paper makes specific, falsifiable numeric claims (e.g., 23% energy reduction, 21% task-failure reduction) that can be checked against the primary literature, but those checks fail without correct references. The paper provides no reproducible code or datasets; its strength is the breadth of topics covered rather than technical depth.","major_comments":[{"comment":"References [3] and [4] are swapped. The MDP-based energy-harvesting dynamic computation offloading work is [4] (Mao, Zhang, and Letaief), but the text attributes it to [3] (Chiang and Zhang, 'Fog and IoT'). Conversely, the Fog and IoT overview is [3], but it appears as [4] in the paragraph on fog computing. The 17% task-completion-time improvement claim in Section III.B cites [1][3]; it should cite [1][4] (or [1] alone if the comparison is internal to the DROO paper). As written, a reader cannot verify the comparison.","section":"III.A and III.B"},{"comment":"The claim that 'DROO exhibited a lower task failure rate, averaging less than 5% compared to the 12% failure rate seen with heuristic-based methods [1][2]' cites [2], which is 'Wireless Powered Communication: Opportunities and Challenges,' a paper that does not present a heuristic offloading baseline. The baseline is presumably reported in [1] (the DROO paper) or needs a different supporting reference. This citation error directly affects the reliability of the stated task-failure comparison.","section":"III.B"},{"comment":"The 17% load-balancing improvement from the DDQN-based traffic-steering xApp is attributed to reference [7] (the X5g testbed paper), but the actual traffic-steering work is reference [10] (Lacava et al., 'Programmable and Customized Intelligence for Traffic Steering in 5G Networks Using Open RAN Architectures'). The entire traffic-steering paragraph is meant to summarize [10], so this mismatch is load-bearing for the self-healing section's quantitative evidence.","section":"IV.B"},{"comment":"The 21% task-failure reduction from DRL-based xApps in ColO-RAN is attributed to reference [9] (O-RAN Performance Analyzer), but the described ColO-RAN experimentation is reference [8] (Bonati et al.). Conversely, the 22% congestion-event reduction is attributed to [8], but the O-RAN Performance Analyzer is [9]. These two citations are cross-swapped, undermining the verifiability of two separate quantitative claims in the same paragraph.","section":"IV.B"}],"minor_comments":[{"comment":"The manuscript self-identifies as a 'reading report,' which is informal for a journal. Please reframe the paper as a survey/tutorial and define the scope and selection criteria for included works.","section":"Abstract and I"},{"comment":"The reference list format is inconsistent: some entries include URLs and arXiv identifiers ([7], [9], [10]) while others provide only bibliographic data. Standardize all entries to a single citation style.","section":"References"},{"comment":"The text states that NVIDIA ARC has validated AI-based mechanisms but provides no citation for this claim; either add a supporting reference or remove the specific mention.","section":"IV.B"},{"comment":"The naming of ColO-RAN is inconsistent (ColO-RAN, Colo-RAN, COLO-RAN) and 'xApp' is not defined at first use. Please standardize terminology and define acronyms.","section":"Throughout"},{"comment":"The phrase 'the heuristic approaches presented in the paper \"Wireless Powered Communication: Opportunities and Challenges\"' conflates the heuristic baselines of [1] with a paper that does not discuss heuristic offloading; rephrase to distinguish the cited references from the baseline methods of [1].","section":"III.B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as a class project and would need substantial editorial work to meet journal standards. The citation cross-swaps are systematic and suggest the reference list was renumbered carelessly. I recommend asking the authors to verify every numerical claim against its stated source and to provide a table mapping each reported percentage improvement to the specific experiment and reference. After these corrections the paper could be a serviceable survey, but in its current form the central quantitative assertions are unverifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a reading report, not a research paper. The authors say so themselves, both in the abstract and in Section II. So the right question is not whether the results are new, but whether the summary is accurate and useful. On that, I have mixed feelings.\n\nWhat the paper does well: it gives a clean tour of two important areas—DROO for MEC offloading and DRL-based xApps in ORAN—and it spells out real limitations (delayed feedback, exploration-exploitation, fault tolerance) that a newcomer would want to see. The structure is sensible, and the choice of key papers (Huang et al., ColO-RAN, the traffic-steering work) is defensible.\n\nThe soft spot is real and load-bearing: the quantified performance claims are attached to the wrong references in several places. In Section III.B, the 17% task-completion-time improvement attributes the MDP energy-harvesting baseline to [1][3], but [3] is the fog/IoT paper; the actual MDP paper is [4]. The <5% vs 12% task-failure comparison cites [1][2], and [2] is not a heuristic offloading baseline. In Section IV.B, the traffic-steering load-balancing figure is cited to [7] instead of [10], the ColO-RAN task-failure number to [9] instead of [8], and the congestion figure to [8] instead of [9]. Because these percentages are the only quantitative evidence for the paper's central \"AI/ML helps\" claim, the mismatch means a reader cannot verify the claim from the cited sources without going to the original papers and checking. That is more than a typo; it is a reproducibility problem in a review.\n\nI do not think this is dishonest. The authors seem to have summarized from memory and made attribution errors. But it does mean the report's evidence is hearsay until fixed.\n\nWho is this for? A newcomer to MEC/ORAN who wants a quick map of the literature might get some orientation. A researcher would not cite it for any number. It does not deserve peer review as a research contribution. If it is meant as a survey for a venue like IEEE Communications Surveys & Tutorials, it would need a full rewrite with accurate citations and more systematic coverage.\n\nRecommendation: desk reject for a research track; if it comes to you as a survey, send it back for major revision on citations before any reviewer sees it.","headline":"A transparent, clearly-labeled reading report with no new results and several mismatched citations that undermine its only quantitative evidence.","tokens_in":8432,"tokens_out":2603,"would_cite":false,"duration_ms":25603,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This review paper argues that deep reinforcement learning and related AI/ML methods deliver concrete, quantifiable gains in mobile edge computing and Open RAN, reporting improvements such as 23% lower energy use and 21% fewer task failures.","keywords":["mobile edge computing","Open RAN","deep reinforcement learning","computation offloading","xApps","network slicing","self-healing networks","next-generation wireless"],"falsifier":"Run DROO in a live, wireless-powered MEC testbed under the same conditions as the original study, measuring energy consumption and task failure rate against the same heuristic and MDP baselines; if the 23% energy reduction and sub-5% failure rate do not reproduce, the paper's central quantitative claim fails. Similarly, deploy the DDQN traffic-steering xApp on an ORAN testbed and compare load-balancing metrics against a static baseline to test the 17% figure.","tokens_in":7374,"feed_emoji":"📡","tokens_out":7402,"duration_ms":60129,"temperature":0.7,"pith_summary":"This is a review paper that argues AI and machine learning, especially deep reinforcement learning (DRL), are the key enablers for mobile edge computing (MEC) and Open RAN (ORAN) in next-generation wireless networks. It surveys the literature and highlights concrete performance numbers: DROO, a DRL-based offloading algorithm, cuts energy consumption by 23% versus heuristics, and DRL-based xApps in the ColO-RAN testbed reduce task failures by 21%. The paper frames these gains as evidence that closed-loop, AI-driven control can make networks more adaptive, energy-efficient, and self-healing. A sympathetic reader would take away that DRL has moved from theory to validated testbed results in both domains.","feed_headline":"DRL offloading cuts energy 23%, xApps cut failures 21%","feed_subtitle":"A review argues AI-driven closed-loop control delivers these gains in next-gen wireless networks.","key_machinery":"The mechanism carrying the argument is the closed-loop control loop enabled by the near-real-time RAN Intelligent Controller (RIC) and its xApps in ORAN, paired with the DRL-based offloading policy in MEC. The DROO algorithm uses deep reinforcement learning to map observed network states to offloading decisions without an explicit model of the system, letting it adapt to changing user demand and channel conditions. In ORAN, xApps hosting DDQN, SVM, k-NN, and decision-tree models run inside the near-real-time RIC to steer traffic, allocate resources, predict interference, and preempt congestion. These mechanisms are what turn network optimization into a learnable, adaptive decision problem, and they are what the paper's reported performance gains are attributed to.","core_discovery":"The central claim is that AI/ML, and deep reinforcement learning in particular, transforms MEC and ORAN from static, model-dependent systems into adaptive, learning-driven ones. In MEC, the DROO algorithm learns optimal computation-offloading decisions for wireless-powered networks, reporting a 23% reduction in energy consumption relative to heuristic approaches, a 17% reduction in average task completion time relative to an MDP-based energy-harvesting method, and a task failure rate below 5% versus 12% for heuristics. In ORAN, the ColO-RAN testbed integrates DRL-based xApps on the near-real-time RIC and achieves a 21% reduction in task failures; other ML-based xApps are reported to cut interference by 25%, improve load balancing by 17%, improve handover success by 18%, and reduce congestion events by 22%. The paper presents these numbers as evidence that closed-loop, AI-driven control is the path to low-latency, reliable, self-healing next-generation radio networks.","pith_inferences":["The reported gains come from specific testbeds and simulations; the paper does not establish that they survive in commercial deployments with non-stationary traffic, hardware heterogeneity, and imperfect telemetry.","An implicit implication is that MEC and ORAN are complementary halves of the same AI control problem: offloading decisions at the edge and radio resource decisions at the RAN both depend on latency, load, and channel state, so joint optimization could compound the reported gains.","A testable extension would be to benchmark DROO not just against the older heuristic and MDP baselines in the paper, but against modern offline-optimal and model-predictive offloading schemes in a live edge testbed, to see whether the 23% energy gain persists.","The paper's suggestion to add delay-tolerant and federated learning mechanisms to DROO implies that its current performance depends on continuous, low-latency feedback; how much of the 23% gain would degrade under intermittent connectivity is left unquantified."],"forward_implications":["DRL-based offloading could let mobile devices in wireless-powered MEC networks make near-optimal offloading decisions without knowing the system dynamics, saving energy and reducing task failures in practice.","ORAN operators could deploy self-healing xApps that detect and mitigate congestion, interference, and handover failures in real time, improving QoS for video streaming, gaming, and VR.","The near-real-time RIC becomes the central decision-making hub, enabling closed-loop control that adapts to changing traffic and radio conditions within milliseconds.","The cited numbers provide concrete, quantifiable targets for validating AI-driven RAN management in experimental testbeds and future field trials.","Combining DRL in MEC with DRL in ORAN could lead to end-to-end optimization across the edge and radio access network, though the paper does not demonstrate this integration."],"supporting_citations":[{"why":"Supplies the DROO algorithm and the central MEC results: 23% energy reduction, 17% shorter task completion time, sub-5% task failure rate.","marker":"[1]"},{"why":"Provides the heuristic wireless-powered communication baseline that DROO is compared against for energy and task-failure results.","marker":"[2]"},{"why":"Provides the MDP-based dynamic offloading approach with energy harvesting that serves as the comparison for the 17% task-completion-time result.","marker":"[4]"},{"why":"Source for the ORAN architecture and the SVM interference-mitigation and k-NN handover results (25% and 18% improvements).","marker":"[5]"},{"why":"Describes the ColO-RAN testbed and the DRL-based xApps credited with a 21% reduction in task failures.","marker":"[8]"},{"why":"Source for the O-RAN performance analyzer and its 22% reduction in congestion events.","marker":"[9]"},{"why":"Cited for the DDQN-based traffic-steering xApp and its reported 17% improvement in load balancing.","marker":"[7]"}],"fun_headline_variants":["AI-driven MEC and ORAN: 23% energy cut, 21% fewer failures","Deep RL cuts MEC energy 23%, xApps cut ORAN failures 21%","Learning wireless networks: 23% less energy, 21% fewer failures","AI transforms MEC and ORAN: 23% energy savings, 21% fewer failures","In MEC and ORAN, AI cuts energy 23% and failures 21%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the performance percentages quoted from the cited papers are accurate, correctly attributed, and will transfer from the specific testbeds and simulations to production networks.","fun_headline_variants_meta":{"raw":{"variants":["AI-driven MEC and ORAN: 23% energy cut, 21% fewer failures","Deep RL cuts MEC energy 23%, xApps cut ORAN failures 21%","Learning wireless networks: 23% less energy, 21% fewer failures","AI transforms MEC and ORAN: 23% energy savings, 21% fewer failures","In MEC and ORAN, AI cuts energy 23% and failures 21%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000659,"raw_usage":{"total_tokens":3005,"prompt_tokens":925,"completion_tokens":2080,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":1965}},"tokens_in":541,"tokens_out":2080,"duration_ms":14359,"temperature":1.0,"reasoning_tokens":1965,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T10:45:15.936435+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DROO in a live, wireless-powered MEC testbed under the same conditions as the original study, measuring energy consumption and task failure rate against the same heuristic and MDP baselines; if the 23% energy reduction and sub-5% failure rate do not reproduce, the paper's central quantitative claim fails. Similarly, deploy the DDQN traffic-steering xApp on an ORAN testbed and compare load-balancing metrics against a static baseline to test the 17% figure.","supporting_citations":[{"cited_title":"Huang, S","cited_arxiv_id":null,"evidence_quote":"Supplies the DROO algorithm and the central MEC results: 23% energy reduction, 17% shorter task completion time, sub-5% task failure rate."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the heuristic wireless-powered communication baseline that DROO is compared against for energy and task-failure results."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MDP-based dynamic offloading approach with energy harvesting that serves as the comparison for the 17% task-completion-time result."},{"cited_title":"Polese, L","cited_arxiv_id":null,"evidence_quote":"Source for the ORAN architecture and the SVM interference-mitigation and k-NN handover results (25% and 18% improvements)."},{"cited_title":"Bonati, S","cited_arxiv_id":null,"evidence_quote":"Describes the ColO-RAN testbed and the DRL-based xApps credited with a 21% reduction in task failures."},{"cited_title":"Kouchaki, S","cited_arxiv_id":null,"evidence_quote":"Source for the O-RAN performance analyzer and its 22% reduction in congestion events."}],"review_version":1}