{"id":"246e887f-7973-4ed4-b894-e0cb2be539e3","arxiv_id":"2507.09124","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CAORA is an O-RAN-based orchestrator combining LSTM traffic forecasts with a Soft Actor-Critic agent to dynamically share GPU instances between RAN and AI workloads, achieving about 90% combined demand fulfillment in trace-driven simulations.","lead":"This paper proposes CAORA, an O-RAN-based architecture that runs a reinforcement learning orchestrator to split GPU resources between real-time 5G radio workloads and AI workloads. Trace-driven simulations on Barcelona 5G data report about 90% combined demand fulfillment, but the results rely on a simplified synthetic AI demand model and no comparison to other learning-based schedulers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (12) maps RNTI counts to MIG demand by min-max normalization, forcing the trace peak to equal exactly 21 MIGs; the 'near 99% fulfillment' claim is thus an artifact of an unvalidated scaling unless a physical RNTI-to-GPU mapping is provided.","rationale":"The reader's weakest assumption points to exactly this issue: the min-max normalized RNTI count is used as a proxy for GPU/MIG demand without validation, and the scale of the proxy controls the headline numbers. I agree this is the single most load-bearing concern. The paper is internally coherent as an architecture proposal and the trace-driven simulation is a genuine independent evaluation effort, so the concern is not fatal by itself. However, all of the quantitative evidence for 'near 99% fulfillment' and 'maximized utilization' flows through Eq. (12). If that mapping is wrong, the comparison between CAORA and the static baselines changes, and the central claim is no longer supported. A measured calibration or an explicit scaling sweep would settle the issue. Other weaknesses (synthetic AI demand, no error bars, missing learning baselines, equation inconsistencies) are real but secondary: they affect generalizability and presentation, not the validity of the main numerical claims. Since my concern is the same one the reader identified and the appropriate remedy is the same set of conditions, I would keep the CONDITIONAL verdict rather than strengthen it to REJECT or weaken it to ACCEPT.","tokens_in":21279,"tokens_out":5044,"duration_ms":66178,"concrete_test":"Re-run the LCCM/EB/PS simulations with Eq. (12) replaced by d_RAN(t) = c * (RNTI_t - min(RNTI)) / (max(RNTI) - min(RNTI) + eps) for c in {0.5, 1.0, 1.5}, keeping the same LSTM-SAC policy and both static baselines. If RAN completion remains near 99% and AI completion stays in the reported ranges across all c, the normalization is not load-bearing. If RAN completion drops by more than ~5 points for c = 1.5 or AI completion collapses for c = 0.5, the headline claims are artifacts of the min-max scaling. Ideally, replace min-max scaling entirely with a physical calibration, e.g., MIGs per active user derived from O-DU profiling in OpenAirInterface or NVIDIA ARC, and report how much of the claimed benefit survives under that calibrated mapping.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim ('near 99% fulfillment') rests on Eq. (12): d_RAN(t) = (RNTI_t - min(RNTI)) / (max(RNTI) - min(RNTI) + eps), and the paper then reads the result as the fraction of Rmax = 21 MIGs required by RAN. This makes the peak RNTI in each trace coincide, by construction, with the entire 21-MIG capacity. No evidence is offered that RNTI counts (numbers of active user identifiers) map linearly, or at all, to GPU/MIG compute demand in an O-DU. Completion rate therefore measures how well the scheduler handles a capacity ceiling placed exactly at the trace maximum; it is not a measured property of an actual AI-RAN system. The same normalization also shapes the baselines: Balanced (50%) and RAN-Priority (70%) are guaranteed to clip exactly at RNTI peaks, and the amount of clipping is set by the arbitrary choice of denominator. Since the optimization in Eqs. (6)-(8) and the reward in Eq. (10) all consume the same normalized demand, the entire 'proactive vs static' gap may be a consequence of the scaling rather than of LSTM forecasting or SAC adaptation. Section VII-A even states that the scaled values 'are then interpreted as the proportion of maximum available resources required by the demand,' with no calibration or validation step. This is the load-bearing weak point for the headline: if an independently measured RNTI-to-MIG mapping places the peak below or above 21 MIGs, the reported completion rates (95.9-99.3% RAN, 56-90% AI) and the conclusions drawn from them will move.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CAORA, a four-layer O-RAN-compliant architecture for sharing GPU/MIG resources between RAN and AI workloads. The E2E orchestrator consumes radio analytics from NRT-RIC xApps via the Y1 interface, uses a SpikeAwareLSTM for multi-horizon demand forecasting and spike detection, and employs a Soft Actor-Critic (SAC) agent to adjust MIG allocations under rate and capacity constraints. The authors evaluate the framework in trace-driven simulations using Barcelona 5G RNTI traces from three areas (Les Corts-Camp Nou, Poble Sec, El Born) plus a synthetic periodic AI-demand model, comparing against balanced (50/50) and RAN-priority (70/30) static baselines on completion rate and adaptability. They report RAN completion rates of 95.9-99.3% and AI completion rates of about 81-90%, concluding that predictive LSTM-SAC orchestration is a viable blueprint for AI-RAN converged 6G systems.","tokens_in":21657,"tokens_out":7292,"duration_ms":86783,"significance":"The architectural contribution is timely: integrating monitoring xApps with an E2E orchestrator over the Y1 interface and dynamically partitioning MIG resources between RAN and AI workloads directly addresses an open problem in AI-RAN convergence. The SpikeAwareLSTM multi-task design is a reasonable mechanism for coupling forecasting with spike detection. If the demand model were independently calibrated, the paper would offer a useful blueprint for future 6G resource management. However, the central quantitative claim rests on an unvalidated RNTI-to-MIG scaling, and the evaluation lacks the ablations and statistical reporting needed to attribute the observed gains to proactive forecasting or SAC adaptation. The work is therefore of moderate-to-good significance but currently does not support 'near 99% fulfillment' as a measured system-level property.","major_comments":[{"comment":"The RAN demand proxy d_RAN(t) = (RNTI_t - min(RNTI))/(max(RNTI) - min(RNTI) + epsilon) is min-max normalized and then interpreted as a fraction of R_max = 21 MIGs. This makes the peak RNTI of each trace coincide exactly with full system capacity by construction. No evidence is provided that RNTI counts (numbers of active user identifiers) map linearly, or at all, to GPU/MIG compute demand of an O-DU. Because the optimization objective, reward, and baselines all consume this same normalized demand, the reported completion rates and the CAORA-vs-baseline gap may be artifacts of the chosen denominator rather than measured properties of the architecture. Please calibrate the mapping on an O-DU/MIG profiling study, provide a sensitivity analysis over plausible scaling laws, and state explicitly that the completion-rate numbers are conditional on this proxy.","section":"Section VII-A, Eq. (12)"},{"comment":"The reward definitions are inconsistent. Eq. (10) defines R'_t as sum_x [C_x(t)/d_x(t) + mu * C_hat_x(t+1)/d_x(t)] minus penalty terms, with C_hat_x(t+1) never defined, while Algorithm 1 uses p_x(t)*C_x(t)/d_x(t) + mu * min(p_x(t)*r_x(t), d_hat_x(t+1))/d_x(t). The priority weighting p_x appears only in the pseudocode, and C_hat_x never appears in the state or allocation equations. Since the SAC agent is trained on this reward, the discrepancy makes the actual optimized objective ambiguous and affects reproducibility.","section":"Section IV-D2, Eq. (10); Algorithm 1, line 10"},{"comment":"The attribution of the performance gain to proactive LSTM forecasting and SAC adaptation is not supported by the present experiment set. The baselines are static (50/50 and 70/30) with no learning or lookahead, and there is no ablation comparing SAC with and without the LSTM forecasts, or against a reactive adaptive policy. Add such ablations, or soften the conclusion that 'predictive orchestration significantly improves' performance.","section":"Section VII-C.2, Figures 9-11"},{"comment":"The evaluation reports only point estimates. No error bars, confidence intervals, number of random seeds, or variance across training runs are provided, so it is impossible to assess whether the reported differences are statistically significant. The text calls the adaptability result 'statistically significant' without performing or citing a significance test. Add repeated-run statistics and a significance test for the completion-rate and adaptability comparisons.","section":"Section VII-B, Evaluation Metrics; Figures 9-12"},{"comment":"Several load-bearing quantities are under-specified. In Eq. (4), Delta(k,tau) is used but never defined, and the relationship between C_pred, Delta_total, r(k,t), and d_x(t) is not established, making it impossible to verify the dimensional consistency of the contention factor and the freed-resource prediction in Eq. (5). In addition, the numerical values of the reward weights mu, lambda, eta, the future discount beta, the rate limits nu_max_x, and the borrowing coefficients alpha_delta in constraint (7) are not reported in Table III or the text. These parameters determine the agent's behavior, so their absence blocks reproduction.","section":"Section IV-B, Eqs. (4)-(5); Section VII-A"}],"minor_comments":[{"comment":"The abstract claims 'near 99% fulfillment of RAN demands,' but the Poble Sec scenario reports a RAN completion rate of 95.9%; please rephrase to report the actual range accurately.","section":"Abstract; Section VII-C.2, Figure 11"},{"comment":"Eq. (13) is not a full-wave rectified sine; it is a shifted sine with a DC offset. Either correct the formula or change the terminology to avoid a misleading description.","section":"Section VII-A, Eq. (13)"},{"comment":"Table III lists R_max = 21 MIGs with the MIG profile 1g.5gb providing 7 MIG instances per A100; clarify whether 21 MIGs corresponds to three GPUs and describe how discrete MIG partitioning is modeled in the simulator.","section":"Table III"},{"comment":"Algorithm 1 outputs {Delta_r_RAN, Delta_r_AI, nu_RAN, nu_AI}, while Section V-A states |A| = 2 and Section VII-A describes actions as {Delta_r_RAN, Delta_r_AI}; align the action-space definition across the paper.","section":"Algorithm 1; Section V-A"},{"comment":"In the submitted PDF, Figures 3-8 contain corrupted font encoding in the axis labels and legends, so the forecast and allocation plots cannot be visually verified; please regenerate the figures.","section":"Figures 3-8"},{"comment":"The text describes the demand fusion as 'Bayesian averaging,' but Eq. (4) uses a simple weighted average of current and predicted demands; rename this to avoid an unjustified Bayesian claim.","section":"Section IV-B, Eq. (4)"},{"comment":"The paper says the LSTM is trained and tested on separate CSV files, but it does not report the exact evaluation day for each trace or confirm that no training data overlaps the test period; provide this information.","section":"Table II; Section VI"},{"comment":"The paper does not state whether code or processed data will be released; given that the raw traces are public (refs. [27], [28]), a reproducibility statement would strengthen the work.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main concern is not the architecture's feasibility but the empirical grounding of the headline number. The RNTI-to-MIG scaling in Eq. (12) is the load-bearing weakness: it forces the peak demand to match the capacity ceiling, so the reported completion rates and the advantage over static baselines may be by construction. If the authors can provide an independent calibration study, or alternatively re-frame the results as conditional on a stated scaling law, and if they add the missing ablations and statistical reporting, the paper could become suitable for publication. I do not see a novelty-disclosure problem; the related work is adequately cited. The missing reward parameter values and the reward inconsistency between Eq. (10) and Algorithm 1 must also be fixed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is CAORA: a Y1-interface monitoring xApp feeding an LSTM+SAC orchestrator that partitions MIGs between RAN and AI on shared GPUs. That specific integration is genuinely new in the O-RAN/AI-RAN space, and the authors use real Barcelona 5G traces with train/test separation, which is more than many simulation papers do. The architectural story is coherent and the figures show the expected behavior.\n\nThe soft spots are in the evaluation. The abstract claims near 99% RAN fulfillment; the reported numbers range from 95.9% to 99.3% depending on scenario, so 'near 99%' overstates the Poble Sec result. More importantly, Eq. (12) min-max normalizes RNTI counts and then reads that as a fraction of 21 MIGs. That forces the peak RNTI to equal full capacity by construction. There is no calibration or physical evidence that RNTI count maps linearly to GPU/MIG demand. The completion-rate metric is also the first term of the reward, so the reported percentages are partly the agent grading itself on what it optimizes. The AI workload is a synthetic sine wave; there are no error bars; the baselines are static (50/50 and 70/30); there's no LSTM ablation; and Eqs. (4), (10), and Algorithm 1 don't fully agree on the reward form. The paper also doesn't clearly differentiate itself from the authors' own prior work [5]. These are fixable with an extra evaluation section, not fundamental conceptual flaws.\n\nThe citation pattern is honest: the authors cite their own prior work and the AI-RAN Alliance report, and the related work is reasonably covered. Nothing in the paper suggests bad faith. My verdict: the architecture is a plausible blueprint and worth refereeing, but the current evidence doesn't support the strong quantitative claims. A serious referee would need the demand mapping validated, variance reported, and a proper ablation. I'd send it to review with a conditional accept expectation.","headline":"CAORA is a plausible integration blueprint for AI-RAN, but the headline completion numbers rest on an unvalidated RNTI-to-MIG scaling and the evaluation needs work.","tokens_in":22213,"tokens_out":1730,"would_cite":false,"duration_ms":20106,"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":"The paper claims that a proactive O-RAN orchestrator can keep RAN demand fulfillment near 99 percent while reclaiming idle GPU capacity for AI workloads.","keywords":["O-RAN","AI-RAN convergence","resource orchestration","Soft Actor-Critic","LSTM forecasting","Multi-Instance GPU","6G networks","trace-driven simulation"],"falsifier":"Re-run the same trace-driven simulator with RAN demand derived from measured GPU/MIG utilization of real distributed-unit signal processing under the same RNTI loads, instead of the normalized RNTI-to-capacity rule of Eq. (12). If RAN completion falls materially below the reported 95.9–98.6 percent, or if peak demand exceeds the 21-MIG capacity, then the near-99 percent claim does not survive.","tokens_in":21065,"feed_emoji":"📡","tokens_out":9366,"duration_ms":101375,"temperature":0.7,"pith_summary":"The paper proposes CAORA, a four-layer O-RAN architecture in which real-time RAN functions and compute-heavy AI workloads share one GPU/MIG pool. Its claim is that a proactive orchestrator—an LSTM forecaster called SpikeAwareLSTM feeding a Soft Actor-Critic (SAC) reinforcement learning agent—can anticipate RAN demand spikes and shift idle capacity to AI tasks without degrading RAN service. Using real 5G traffic traces from three Barcelona districts, the authors report RAN demand fulfillment of 95.9–98.6 percent and AI workload completion of 78.5–81.5 percent, above fixed 50/50 and 70/30 baselines. The reason to care is economic: if CAORA is right, a 6G operator can buy one accelerated pool instead of dedicating underused GPUs to AI and separate capacity to RAN.","feed_headline":"AI-and-RAN orchestrator keeps RAN demand ~99% met","feed_subtitle":"O-RAN architecture with spike-aware forecasting hands idle GPU slices to AI workloads without starving the radio network.","key_machinery":"The load-bearing machinery is the coupling of a SpikeAwareLSTM forecaster with a Soft Actor-Critic (SAC) agent inside an end-to-end orchestrator. SpikeAwareLSTM is a two-layer LSTM trained with a composite loss, mean-squared error for demand prediction plus binary cross-entropy for spike detection, whose internal spike probability $\\hat{s}^{\\mathrm{int}}_t$ down-weights forecasted demand during anomalies. The SAC agent's state includes current and predicted RAN/AI demands and the previous allocation, and its reward combines current QoS with anticipated QoS while imposing a super-linear overprovisioning penalty and a predictive contention penalty. The resource constraint allows reservations up to $R_{\\max} + \\sum_\\delta \\alpha_\\delta \\hat{\\Delta}_{\\mathrm{free}}(t+\\delta)$, so predicted freed resources can be spent before they appear. The framework partitions the physical GPU into Multi-Instance GPU (MIG) slices, each an isolated compute instance, and exchanges KPIs between a monitoring xApp in the near-real-time RIC and the orchestrator over the Y1 interface.","core_discovery":"The paper's central claim is that RAN and AI workloads do not need separate dedicated GPU pools. CAORA ties a monitoring xApp inside the near-real-time RIC to an end-to-end orchestrator over the Y1 interface; the orchestrator's SpikeAwareLSTM forecasts demand and flags spikes, and the SAC agent translates those signals into rate-limited, priority-scaled MIG allocations that first protect RAN and lend spare capacity to AI. In trace-driven simulations over three Barcelona 5G traces, the authors report RAN demand fulfillment of 95.9–98.6 percent and AI workload completion of 78.5–81.5 percent, with both above the fixed 50/50 balanced and 70/30 RAN-priority baselines. The paper presents CAORA as a blueprint for a converged 6G platform in which one shared accelerated pool carries radio and AI work.","pith_inferences":["The RNTI-to-MIG demand model in Eq. (12) is the least supported link in the chain: it scales the busiest observed RNTI count to exactly $R_{\\max}$, so a perfectly accurate forecaster would satisfy demand almost by construction. I would not transfer the 99 percent figure to a deployment without calibrating that mapping against real GPU utilization.","A clean ablation, replacing SpikeAwareLSTM with a vanilla LSTM while keeping the SAC agent unchanged, would isolate whether joint spike detection is what produces the adaptability gain; the paper does not perform that separation.","The simulations cover a single orchestration site under one SAC agent; a multi-cell or multi-RIC extension would face Y1 monitoring latency, distributed state, and competing allocations, none of which the current results address."],"forward_implications":["If CAORA is right, the Y1 interface can carry radio analytics from a near-real-time RIC to an external orchestrator fast enough for closed-loop AI/RAN co-allocation.","Operators could run AI training, inference, or generative-AI workloads on RAN GPU capacity during off-peak hours while keeping RAN demand fulfillment above 95 percent in the tested scenarios.","In the football-match trace, the reported 98.3 percent RAN completion with 81.5 percent AI completion would come from proactive spike-aware allocation rather than from fixed service priorities.","Since the results hold across stadium-spike, nightlife-cyclical, and residential-steady demand traces, the same control loop could serve as a common resource manager for a city-scale converged 6G platform."],"supporting_citations":[{"why":"supplies the O-RAN architecture, interfaces, and functional split that CAORA builds on.","marker":"[1]"},{"why":"defines the AI-RAN convergence vision whose gap the paper addresses.","marker":"[3]"},{"why":"defines MIG partitioning used as the shared resource model.","marker":"[26]"},{"why":"provides the open Barcelona 5G traffic traces used for demand profiles.","marker":"[27]"},{"why":"supplies the trace dataset and the forecasting baseline the spike-aware LSTM is compared with.","marker":"[28]"},{"why":"justifies the synthetic periodic AI workload pattern used to emulate LLM-inference-like demand.","marker":"[29]"}],"fun_headline_variants":["CAORA: AI and RAN share GPU without starving radio","Proactive orchestration keeps RAN at 99% while AI runs","AI and RAN coexist on shared GPU with smart forecasting","O-RAN orchestrator forecasts spikes, lends GPU to AI","6G blueprint: AI and RAN share GPU via proactive orchestration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that RNTI counts (radio network temporary identifiers, a proxy for active users), after being min-max normalized to [0,1], linearly represent the fraction of the 21-MIG GPU pool that the RAN actually needs, with the busiest observed count set equal to full capacity; if that mapping is wrong, the reported fulfillment percentages change.","fun_headline_variants_meta":{"raw":{"variants":["CAORA: AI and RAN share GPU without starving radio","Proactive orchestration keeps RAN at 99% while AI runs","AI and RAN coexist on shared GPU with smart forecasting","O-RAN orchestrator forecasts spikes, lends GPU to AI","6G blueprint: AI and RAN share GPU via proactive orchestration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000629,"raw_usage":{"total_tokens":2925,"prompt_tokens":982,"completion_tokens":1943,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":1854}},"tokens_in":598,"tokens_out":1943,"duration_ms":15322,"temperature":1.0,"reasoning_tokens":1854,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:03:24.093894+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same trace-driven simulator with RAN demand derived from measured GPU/MIG utilization of real distributed-unit signal processing under the same RNTI loads, instead of the normalized RNTI-to-capacity rule of Eq. (12). If RAN completion falls materially below the reported 95.9–98.6 percent, or if peak demand exceeds the 21-MIG capacity, then the near-99 percent claim does not survive.","supporting_citations":[{"cited_title":"Under- standing O-RAN: Architecture, Interfaces, Algorithms, Security, and Research Challenges,","cited_arxiv_id":null,"evidence_quote":"supplies the O-RAN architecture, interfaces, and functional split that CAORA builds on."},{"cited_title":"AI-RAN Alliance Vision and Mission White Paper,","cited_arxiv_id":null,"evidence_quote":"defines the AI-RAN convergence vision whose gap the paper addresses."},{"cited_title":"NVIDIA Multi-Instance GPU User Guide,","cited_arxiv_id":null,"evidence_quote":"defines MIG partitioning used as the shared resource model."},{"cited_title":"Federated-time-series-forecasting,","cited_arxiv_id":null,"evidence_quote":"provides the open Barcelona 5G traffic traces used for demand profiles."},{"cited_title":"Federated Learning for 5G Base Station Traffic Forecasting,","cited_arxiv_id":null,"evidence_quote":"supplies the trace dataset and the forecasting baseline the spike-aware LSTM is compared with."},{"cited_title":"How ACK Edge Solves Challenges in Elasticity for LLM Inference Services,","cited_arxiv_id":null,"evidence_quote":"justifies the synthetic periodic AI workload pattern used to emulate LLM-inference-like demand."}],"review_version":1}