{"id":"761828df-4e3e-4baa-a6ff-b1c75137c06e","arxiv_id":"2509.10097","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A hybrid heuristic and K-Means xApp cuts emulated O-RAN energy use by about 13% by sleeping underutilized micro cells while keeping throughput nearly unchanged.","lead":"The authors build an O-RAN energy saving xApp that switches underused small cells to sleep using simple rules plus K-Means clustering. In a 2-hour emulated urban network with 246 users and 51 cells, it cut power by about 13% while keeping user throughput nearly unchanged.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 13.27% savings does not follow from Table III: (4.87 - 4.32)/4.87 = 11.29%, and the 4.87 kW baseline is also difficult to reconcile with Table I's all-on max of 10.84 kW.","rationale":"The paper's contribution is the measured energy saving: 'approximately 13% energy savings' appears in the abstract, contributions, and conclusion, and Table III is the only quantitative support. A direct arithmetic check of Table III yields 11.29% rather than 13.27%. This is not a matter of interpretation or consensus; it is an internal inconsistency in the paper's central evidence. If the correct number is 11.3%, the headline claim is overstated by about 2 percentage points, which is material for a systems paper whose novelty is largely quantitative. The second inconsistency, the 10.84 kW all-on maximum implied by Table I versus the 4.87 kW baseline, is also concerning but could be explained by load-dependent power consumption: cells at lower utilization draw less than maximum. The paper, however, does not document the power model, so the baseline cannot be independently assessed. I agree with the reader's conditional verdict: the result may well be correct, but the onus is on the authors to provide cell-level power logs or a power-model description. I do not see evidence of fraud or obvious algorithmic error; the issue is verifiability and a concrete numerical discrepancy. The reader's weakest assumption correctly identified the power-model gap, though not the arithmetic mismatch, hence partial agreement.","tokens_in":8574,"tokens_out":3775,"duration_ms":31966,"concrete_test":"Obtain the raw per-cell power traces (or per-cell KPM power fields) from the TeraVM AI RSG logs for the baseline, heuristic, and hybrid runs. Recompute aggregate power and the reduction percentage directly from those traces. If aggregate powers are 4.87 and 4.32 kW, the reported 13.27% must be corrected to 11.29%. If 13.27% is retained, the proposed xApp power must be about 4.22 kW. Also sum logged per-cell powers under the all-on baseline and compare with 10 x 379 + 41 x 172 = 10.84 kW; if the discrepancy persists, the power model used in the emulator must be documented before any absolute savings figure can be cited.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests entirely on the energy numbers in Table III. Two internal inconsistencies make that foundation insecure. First, the arithmetic: (4.87 - 4.32)/4.87 = 0.1129, i.e. 11.29%, not the reported 13.27%. For 13.27%, the proposed run would need to consume about 4.22 kW rather than 4.32 kW. Second, the baseline itself: Table I lists max power consumption as 379 W per MACRO and 172 W per MICRO, so with 10 MACRO and 41 MICRO all active the naive all-on maximum is 10.84 kW, more than double the reported 4.87 kW baseline. The paper never specifies the emulator's power model (load-dependent component, sleep-state accounting, time-averaging over the 2-hour run), so the reader cannot tell whether the baseline is realistic or whether the savings percentage is an artifact of the digital twin's unvalidated energy model. The QoS-preservation claim is secondary: even if it holds, the magnitude of the energy saving is the headline contribution, and that magnitude is not internally consistent as reported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a hybrid Energy Saving (ES) xApp for O-RAN that combines heuristic switch-off rules with K-means-based cell activation, integrated with the VIA VI TeraVM AI RSG digital twin. The emulation covers 51 cells (10 MACRO, 41 MICRO) and 246 UEs in a dense urban scenario over two hours. The reported results in Table III are a baseline of 4.87 kW power / 2.47 Gbps downlink throughput, a heuristic xApp at 4.53 kW / 2.39 Gbps, and the proposed hybrid xApp at 4.32 kW / 2.46 Gbps; the manuscript claims approximately 13% energy savings and only 0.4% throughput degradation. Section IV formulates RU sleep control as a mixed-integer optimization problem, and Section V describes the heuristic and unsupervised-learning components of the proposed xApp.","tokens_in":8799,"tokens_out":5114,"duration_ms":43131,"significance":"The paper addresses a relevant operational problem and uses a commercial-grade emulator with O-RAN interfaces, which is a strength. The proposed design is lightweight and plausibly deployable, and the comparison against a heuristic baseline is useful. However, the central quantitative claim is not self-consistent as reported: the 13.27% saving does not follow from the numbers in Table III, and the baseline power is not reconciled with the per-cell power ratings in Table I. These issues must be resolved before the contribution can be fully assessed. If corrected, the work would be a useful experimental contribution to O-RAN energy-efficiency research.","major_comments":[{"comment":"The reported 13.27% saving for the proposed xApp is inconsistent with the numbers in Table III: (4.87 − 4.32)/4.87 = 11.29%. To obtain 13.27%, the proposed configuration would need to consume about 4.22 kW, not 4.32 kW. This is the headline result, so the table or the text must be corrected, or the discrepancy must be explicitly explained.","section":"Section VI, Table III"},{"comment":"The baseline of 4.87 kW is difficult to reconcile with the per-cell maximum power ratings in Table I: 10 MACRO cells at 379 W plus 41 MICRO cells at 172 W gives an all-on maximum of approximately 10.8 kW, more than twice the reported baseline. No power-model details (load-dependent consumption, time averaging over the two-hour run, sleep-state accounting, or the effect of the 3 dB/s power reduction ramp) are provided, so the reader cannot determine whether 4.87 kW is a realistic baseline. Please specify the digital twin's power model and how the baseline and reported averages are computed.","section":"Section III, Table I and Section VI, Table III"},{"comment":"The paper does not report the number of emulation seeds, confidence intervals, or error bars. Section II states that simulations with different seeds produce different behavior, but Section VI presents only single-point averages. Without this information, the 6.98% versus 13.27% comparison and the 0.4% throughput difference cannot be statistically evaluated. The authors should report means and standard deviations over multiple seeds.","section":"Section VI"},{"comment":"The QoS-preservation claim is supported only by aggregate downlink throughput. The manuscript's own logging submodule records throughput outage and per-cell load metrics, but these are not reported. Equation (6) and Algorithm 1 assume that a neighbor cell with PRB utilization below rho and RSRP above R_min guarantees a successful handover; handover failure and interruption are not modeled. To substantiate the claim of maintaining QoS, the paper should report outage counts, handover-related events, and per-UE throughput distributions, not only aggregate throughput.","section":"Section V.B, Eq. (6), and Section VI"}],"minor_comments":[{"comment":"The text contains several typos, including 'Tabled III' instead of 'Table III', 'assosiated', 'acceptabale', 'attched', and 'it’s vicinity'.","section":"Section VI"},{"comment":"The speed of fast-car UEs is written as '15/m' and should be '15 m/s'.","section":"Table II"},{"comment":"Equation (1c) uses R_min as a minimal acceptable signal power, while Eq. (6) uses R_min as an RSRP threshold in dBm; the units and definitions should be clarified and made consistent.","section":"Section IV and Eq. (6)"},{"comment":"The figures would benefit from labeled axes and explicit legends, and Fig. 3 should state whether the plotted power is instantaneous or time-averaged.","section":"Figures 3 and 4"},{"comment":"Line 12 sets the KMeans number of clusters to k=|C_near_sleep|; if the number of UEs in a cluster is smaller than the number of sleeping cells, the clustering assignment is ill-posed and should be discussed or guarded against.","section":"Section V, Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick take: this is a cleanly written systems paper showing a hybrid heuristic+K-Means xApp on a 51-cell, 246-UE TeraVM AI RSG emulation, with a claimed 13% energy saving. The architecture is sensible, the algorithm is lightweight and explainable, and the comparison against a heuristic-only xApp is a nice touch. The paper earns credit for evaluating at a scale larger than the group's prior work and for describing the digital twin setup in enough detail to reproduce the scenario, though not the tool's internals.\n\nThe soft spots are real. The headline saving does not follow from Table III: (4.87 - 4.32)/4.87 is 11.3%, not 13.27%. That is not a rounding nit; the percentage is off by two points. The baseline power of 4.87 kW also needs a one-line explanation: with 379 W per macro and 172 W per micro, an all-on maximum would be near 10.8 kW, so the emulator clearly uses a load-dependent or time-averaged model that the paper never specifies. Without that, a reader cannot tell if the saving is an artifact of the energy model. There are also no error bars, no seeds, and no indication of run-to-run variance; the Simulation Automation section mentions seed control, but the results section does not use it. No code or data is released, so the emulator output cannot be independently checked.\n\nThe QoS claim (0.4% throughput loss) is secondary and more plausible, but it relies on Eq. (6) conditions implying a clean handover. That is reasonable as a heuristic, not as a guarantee.\n\nNone of this suggests the method is a sham. The algorithm is a modest combination of known building blocks—cell sleep control, K-Means, threshold rules—and the paper says as much. The self-citations are to prior xApp work, which is normal in a research group line. The citation pattern is fine.\n\nWho is this for? People working on O-RAN energy-saving xApps or digital twin evaluation will get a useful reference architecture and a benchmark scenario. The specific numbers need a correction before being quoted.\n\nRecommendation: send it to peer review, but require the authors to fix the arithmetic, explain the power model, and report variance over multiple seeds. It is a legitimate systems contribution that needs a revision, not a desk reject.","headline":"A worthwhile hybrid xApp demo on a large emulated O-RAN, but the headline 13.27% saving contradicts the paper's own Table III arithmetic and needs correction before the number is trusted.","tokens_in":9383,"tokens_out":3614,"would_cite":false,"duration_ms":27408,"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 hybrid xApp that switches off underused micro cells saves 13 percent of RAN energy while keeping downlink throughput nearly unchanged.","keywords":["Open RAN","energy efficiency","xApp","digital twin","sleep mode","unsupervised learning","K-Means clustering","QoS"],"falsifier":"Measure the real power draw of a comparable 51-cell deployment (or a higher-fidelity simulation with explicit handover failure and interruption modeling) with all cells on; if the all-on baseline is not close to the emulator's 4.87 kW, or if any user loses connectivity during cell shutdown, the 13 percent saving and the QoS-preservation claim would not carry over to practice.","tokens_in":8348,"feed_emoji":"🔋","tokens_out":5593,"duration_ms":41428,"temperature":0.7,"pith_summary":"This paper tries to show that a lightweight xApp—a control application running in the O-RAN near-real-time RIC—can cut RAN energy consumption by putting underused micro cells to sleep without hurting users. The proposed hybrid design uses simple rules to decide which cells to switch off and unsupervised K-Means clustering to decide which sleeping cell to wake when a neighborhood gets congested. The authors validate it in a two-hour emulation of a dense urban network with 51 cells and 246 users, reporting about 13 percent energy savings with only 0.4 percent downlink throughput loss. The point of the work is that energy efficiency does not have to be bought at the expense of quality of service, and that the intelligence can be simple enough to run in real time.","feed_headline":"Hybrid xApp cuts RAN energy use 13 percent with no QoS loss","feed_subtitle":"Rule-based shutdown plus K-Means wake-up saves power in a 51-cell digital twin while throughput stays at 99.6 percent.","key_machinery":"The central object is the hybrid Energy-Saving xApp: a near-real-time RIC application whose switch-off logic is rule-based—idle cells, or cells with downlink PRB utilisation below a threshold whose users can all be handed over to neighbors with sufficient RSRP, are put to sleep—while its switch-on logic is an unsupervised K-Means clustering step that groups active users and sleeping cells by coordinates and activates the sleeping cell with minimum throughput-weighted distance to the most demanding users. The digital-twin emulator supplies the live key performance measurements and executes the on/off commands, providing the testbed that makes the end-to-end measurement possible.","core_discovery":"On the paper's own terms, the discovery is that a hybrid heuristic-plus-clustering xApp can manage cell on/off states at scale: in the emulated 51-cell dense-urban Open RAN, average power fell from 4.87 kW (all cells on) to 4.32 kW, a 13.27 percent saving, while downlink throughput dropped only from 2.47 to 2.46 Gbps. The heuristic-only baseline saved 6.98 percent but at a 3.32 percent throughput cost, which the hybrid avoids by waking cells precisely where throughput demand is concentrated. The author's claim is that this demonstrates a practical, near-real-time energy-saving controller for O-RAN that preserves user QoS.","pith_inferences":["If the emulation's power model under-represents the real consumption of micro cells, the absolute savings in kW would change even if the control logic works as described; field measurements of per-cell power are the natural next check.","The activation logic could be extended to predict congestion ahead of time using traffic history, turning the reactive wake-up into a proactive one, which might reduce the small throughput dips seen during demand troughs.","Because the clustering step only uses coordinates and throughput demands, the same xApp design could be applied to macro cells or to heterogeneous deployments with different cell types, though the handover-feasibility check would need re-tuning.","The comparison against a heuristic-only baseline suggests that pure rule-based switching may be too blunt in dense urban settings; a similar hybrid pattern may transfer to other domains where the cost of switching states is asymmetric (fast on, slow off)."],"forward_implications":["Operators can deploy energy-saving xApps that need no labelled training data or offline training phase, since the ML component is unsupervised and the rules are lightweight.","The heuristic component can serve as a bounded, predictable fallback if the ML component makes an unreliable decision in a live network.","The same two-stage pattern—rule-based deactivation plus clustering-based activation—could be applied to other RIC use cases such as load balancing or traffic steering.","The near-baseline throughput result suggests that energy savings in O-RAN can be achieved without user-visible QoS degradation, at least under the emulated dense-urban traffic profile.","The architecture is compatible with standard O-RAN interfaces (E2SM-KPM for measurements, E2SM-RC for control), so it can be tested in other RIC-compliant environments."],"supporting_citations":[{"why":"Supplies the digital-twin RAN emulator that generates all power and throughput measurements, including the 4.87 kW baseline and 4.32 kW result.","marker":"[16]"},{"why":"Earlier DQN-based xApp for O-RAN energy saving that this work scales up to a larger emulated scenario.","marker":"[17]"},{"why":"Earlier intelligent xApp deployment that motivates the hybrid design and provides the comparative context.","marker":"[18]"},{"why":"Defines the O-RAN architecture in which xApps and the near-real-time RIC operate, grounding the interfaces used.","marker":"[3]"},{"why":"Specifies the near-RT RIC and E2 interface that carry the KPM reports and cell on/off commands in the design.","marker":"[7]"}],"fun_headline_variants":["Hybrid xApp trims RAN power 13% while keeping QoS","AI-based xApp saves 13% energy in Open RAN test","Smart sleep modes cut RAN energy by 13% without QoS dip","Clustering xApp yields 13% energy cut in O-RAN emulation","Digital twin guides xApp to 13% power savings in RAN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The savings figure assumes the digital twin's power, radio, propagation, and mobility models behave like a real network, and that every user attached to a cell slated for shutdown can actually be handed off before the cell powers down.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid xApp trims RAN power 13% while keeping QoS","AI-based xApp saves 13% energy in Open RAN test","Smart sleep modes cut RAN energy by 13% without QoS dip","Clustering xApp yields 13% energy cut in O-RAN emulation","Digital twin guides xApp to 13% power savings in RAN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000555,"raw_usage":{"total_tokens":2628,"prompt_tokens":915,"completion_tokens":1713,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":1614}},"tokens_in":531,"tokens_out":1713,"duration_ms":383455,"temperature":1.0,"reasoning_tokens":1614,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:57:09.987615+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the real power draw of a comparable 51-cell deployment (or a higher-fidelity simulation with explicit handover failure and interruption modeling) with all cells on; if the all-on baseline is not close to the emulator's 4.87 kW, or if any user loses connectivity during cell shutdown, the 13 percent saving and the QoS-preservation claim would not carry over to practice.","supporting_citations":[{"cited_title":"TeraVM RIC Test,","cited_arxiv_id":null,"evidence_quote":"Supplies the digital-twin RAN emulator that generates all power and throughput measurements, including the 4.87 kW baseline and 4.32 kW result."},{"cited_title":"Energy Saving in 6G O-RAN Using DQN-based xApp,","cited_arxiv_id":null,"evidence_quote":"Earlier DQN-based xApp for O-RAN energy saving that this work scales up to a larger emulated scenario."},{"cited_title":"Enhancing energy efficiency in o-ran through intelligent xapps deployment,","cited_arxiv_id":null,"evidence_quote":"Earlier intelligent xApp deployment that motivates the hybrid design and provides the comparative context."},{"cited_title":"Intelli- gence and learning in o-ran for data-driven nextg cellular networks,","cited_arxiv_id":null,"evidence_quote":"Defines the O-RAN architecture in which xApps and the near-real-time RIC operate, grounding the interfaces used."},{"cited_title":"O-RAN near-RT RIC architecture 4.0,","cited_arxiv_id":null,"evidence_quote":"Specifies the near-RT RIC and E2 interface that carry the KPM reports and cell on/off commands in the design."}],"review_version":1}