{"id":"04955b3f-2db8-42b2-82a5-32d838174005","arxiv_id":"2607.21963","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An O-RAN framework predicts mmWave channel state from car motion messages, with LSTM NMSE ≈0.01 on a simulated scenario and a federated variant reaching 0.0165.","lead":"This paper builds an O-RAN framework that predicts 5G mmWave channel state from vehicle motion messages using lightweight neural networks trained locally or by federated learning. It reports low prediction error on a simulated urban scenario, but temporal data leakage and missing baselines make the accuracy claims uncertain.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-sample min-max normalization (Eq. 3) makes predicted CSI non-invertible and phase-distorted; reported NMSE does not demonstrate usable CSI prediction for beam management.","rationale":"The Reader's rationale already flags the per-sample min-max normalization and the random split, but their stated weakest assumption is the C-V2X CAM overhearing reliability. I agree that CAM loss/delay is a real deployment concern, but it is external to the algorithm: even with perfect CAM reception, the model output is not physically meaningful CSI because of the normalization. This is an internal, mathematically demonstrable flaw that directly undermines the central claim of 'accurate CSI prediction' for beam management. The random 70/30 split is a second serious issue—it makes the reported NMSE optimistic—but it is less fundamental than the normalization. The paper never explains how the normalized prediction is converted back to CSI for DL-BMM; without such a de-normalization step, the framework cannot function as described. Therefore the current manuscript does not support the strong quantitative claim; it requires a major redesign of the prediction target or a fixed normalization scheme, so the verdict should move from CONDITIONAL to REJECT.","tokens_in":19021,"tokens_out":6942,"duration_ms":80621,"concrete_test":"Retrain the LSTM (Section IV-A) with a fixed affine normalization computed once from the training set (or no normalization), and evaluate on a temporally contiguous test set. Then compute both the NMSE and the beam-selection accuracy by feeding predicted CSI into the DL-BMM and comparing the chosen beam with that from the true CSI. If the per-sample-normalized model cannot be de-normalized to produce correct beams without oracle (ground-truth) min/max, the framework's output is not deployable for proactive beam switching.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim requires that the CSI-PM output can drive proactive beam switching. But Eq. (3) normalizes each target CSI vector by its own per-sample min and max, an affine transform h' = (h - c)/s with sample-dependent complex shift c and real scale s. At inference, the model outputs h' only; recovering the physical CSI needs the future sample's min/max, which is exactly the unknown being predicted. The transform also does not preserve the spatial signature: subtracting c and dividing by s changes relative phases/magnitudes across antennas (e.g., [1,-1] maps to [1,0]), which is fatal for beam selection that relies on the array response. Tables V-VII report NMSE in this normalized domain, so the numbers are not evidence of accurate CSI prediction in any operational sense. The random 70/30 split further inflates results by leaking adjacent temporally correlated samples, but even with a clean split the normalization flaw invalidates the quantitative claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an O-RAN-compliant framework for autonomous CSI prediction in 5G mmWave vehicular networks. A gNB collects time-aligned C-V2X CAM mobility data (position, speed, acceleration) and CSI feedback, then trains a lightweight LSTM predictor (with 1D-CNN and Transformer baselines) to output the next CSI vector from a 10-step CAM history. The framework is evaluated per-gNB and in a federated learning mode using the DeepMIMO Dynamic Doppler scenario with 5000 samples at 100 ms periodicity, four gNBs, a 4x4 BS array, and a single-antenna UE for the baseline. Reported NMSE values are 0.0094-0.018 per gNB and 0.0165 for the federated global model, along with complexity, latency, robustness, prediction-horizon, and multi-antenna scalability analyses.","tokens_in":19272,"tokens_out":8823,"duration_ms":98505,"significance":"If the reported accuracy were meaningful in a physical CSI sense, the framework would be a useful step toward proactive beam management with low-overhead, context-aware prediction, and the O-RAN MLOps discussion is a genuine strength. The paper also offers useful architecture comparisons, an FL implementation, and complexity/latency tables. The FL experiment using Flower and the explicit complexity models are transparent. However, the quantitative claims rest on an evaluation protocol that currently does not establish that the predicted quantities are usable physical CSI forecasts or that the model generalizes beyond interpolated training trajectories. The framework idea is sound and the issues are fixable, but the evaluation must be substantially reworked.","major_comments":[{"comment":"The per-sample min-max normalization makes the prediction target non-invertible. The model is trained to predict h_norm = (h - min(h))/(max(h)-min(h)), where min(h) and max(h) are sample-specific statistics of the future CSI vector. At inference, recovering physical h from h_norm requires knowing the future sample's min and max, which are exactly the unknown quantities. The reported NMSE in Eq. (4) and Tables V-VII is therefore computed in the normalized domain. In addition, subtracting the same scalar from both real and imaginary parts of every coefficient is not a complex rotation or scaling, so it changes relative phases and magnitudes across antennas; beam selection depends on the array response. Consequently, even low NMSE in this normalized space does not demonstrate that the output can drive beam switching. Please re-evaluate using a global/fixed normalization or predict the denor","section":"Section IV-A, Eq. (3)"},{"comment":"The random 70/30 split of 5000 temporally adjacent samples, combined with the L=10 sliding-window construction, creates severe temporal leakage. A test sample at time t uses inputs at t-10,...,t-1; under a random split, many of those neighboring samples (and their targets) are also in the training set with high probability. Since DeepMIMO channel coefficients are deterministic functions of vehicle position (Eq. (1)), the model can effectively interpolate the CAM-to-CSI mapping for the same trajectories rather than forecast unseen channel evolution. This likely inflates the NMSE numbers in Table V and also affects the FL evaluation in Section IV-B. Please use a strict temporal split (e.g., a contiguous training prefix and test suffix, or held-out trajectories) and report results for vehicles and time intervals not seen during training.","section":"Section IV-A, train/test split"},{"comment":"The target CSI vector is underspecified with respect to subcarriers. Section II-B defines the full channel H in C^{M x K} with K=240 OFDM subcarriers, but the model output is described as 2M=32 real values for the baseline N=1 UE configuration. It is unclear whether the predictor targets one representative subcarrier, a subset, or a compressed representation, and whether the same target is used for all K subcarriers. This ambiguity changes the meaning of the reported NMSE and affects reproducibility. Please clarify exactly which subcarriers form the prediction target and how the K dimension is handled.","section":"Section II-B, Table I, Section IV-A"},{"comment":"The framework assumes the gNB overhears all CAM messages from surrounding vehicles and that CAM and CSI streams are perfectly time-aligned; Section III states that exact timing/delay analysis is out of scope. Because CAM loss, delay, or missing broadcasts affect both training labels and inference inputs, the reported accuracy presupposes an idealized C-V2X control plane. The abstract and conclusions claim realistic autonomous operation, so this assumption should either be relaxed in the claims or tested via a sensitivity analysis with respect to CAM drop rate and delivery delay.","section":"Section II-A and Section III"}],"minor_comments":[{"comment":"Equation (3) is missing parentheses; it should read h_norm = (h - min(h))/(max(h)-min(h)).","section":"Eq. (3)"},{"comment":"Section II-C says the dataset contains 2000 temporal scenes, while Section IV-A and the FL section state 5000 samples. Please correct the inconsistency.","section":"Section II-C vs Section IV-A"},{"comment":"The first sentence refers to 'architectures introduced in Section IV-B'; this should be Section IV-A.","section":"Section IV-B"},{"comment":"Please state whether the same trained weights are used for all prediction horizons or the model is retrained for each horizon. If the same weights are used, clarify how the input window is adjusted for longer horizons.","section":"Table VI"},{"comment":"The axis label says 'Noise added during testing', but the figure compares three scenarios with different training conditions. Clarify the label or caption to avoid confusion.","section":"Fig. 6"},{"comment":"The DeepMIMO DD1 scenario citation should include the exact version/release and generation parameters to support reproducibility.","section":"Reference [39]"}],"recommendation":"major_revision","confidential_remarks":"The paper has a useful systems-level contribution, but the evaluation must be redone. The normalization issue is the most serious: the reported NMSE is not in a domain that supports the operational beam-management claim. The temporal split issue and subcarrier ambiguity are also load-bearing. I am not recommending rejection because the O-RAN framework, model comparison, and FL implementation are valuable and the flaws are correctable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the framework and the O-RAN MLOps story are the real contribution; the accuracy numbers are not. The per-sample min-max normalization in Eq. (3) quietly changes the channel vector in a way that breaks the link to beam management, and the random 70/30 split leaks temporal correlation. Both are fixable, but the current version overclaims.\n\nWhat's genuinely new: the CAM-only input design — predicting future CSI from position/speed/acceleration without historical CSI — is a clean idea, and the evaluation on the DeepMIMO DD1 scenario with four gNBs plus an FL variant is a concrete step beyond the authors' conference paper. The MLOps lifecycle discussion and the complexity/latency comparison among LSTM, 1D-CNN, and Transformer are useful for anyone building an O-RAN xApp for beam management. The robustness study with noisy CSI and the multi-antenna scalability analysis are nice touches.\n\nThe trouble is the evaluation protocol. Eq. (3) normalizes each CSI vector by its own min and max over the 32 real/imaginary components. At inference, the model outputs the normalized vector, and you cannot invert it without the future sample's min/max — which is the thing being predicted. Worse, subtracting a real per-sample constant and dividing by a real scale does not preserve the array response: relative phases across antennas change, so the predicted vectors would mislead a beam selection algorithm even if the NMSE in the normalized domain looked great. The plotted examples (Figs. 4–5) are all in the normalized [0,1] box, so the NMSE values in Tables V–VII are evidence of fitting normalized shapes, not of usable CSI prediction.\n\nThe random split compounds this. With 5000 consecutive 100 ms samples and a random 70/30 split, the test set is filled with near-duplicates of training samples; the model is interpolating positions it has basically seen. A temporally contiguous split (e.g., train on first 70% of the trace, test on the last 30%) would give an honest picture.\n\nThe CAM-overhearing assumption is explicit but unvalidated — no loss or delay model — though the paper honestly says timing analysis is out of scope. That's a lesser concern than the normalization issue.\n\nWho's this for? Researchers working on O-RAN xApps or CAM-assisted beam management will find the framework discussion worth reading. But the quantitative claims need a redo before they can be cited. I'd send it to a serious referee, with the expectation of major revision.","headline":"The system design and O-RAN MLOps framing are worth a serious look, but the reported NMSE is computed on per-sample normalized CSI that cannot be inverted for beam management, and the random temporal split leaks correlation, so the headline accuracy claim does not hold as stated.","tokens_in":19805,"tokens_out":3413,"would_cite":false,"duration_ms":38780,"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":"This paper argues that 5G base stations can predict future mmWave channel states from nothing but the short-range vehicle status messages they already overhear, making proactive beam switching possible without extra feedback.","keywords":["CSI prediction","mmWave vehicular networks","C-V2X","cooperative awareness messages","LSTM","federated learning","O-RAN","proactive beam switching"],"falsifier":"Take the same urban scenario and run the predictor with a percentage of CAM messages randomly dropped or delayed by more than 100 ms; if NMSE degrades sharply (e.g., doubles) under realistic loss rates, the autonomy claim fails in practice. More directly, a synthetic dataset in which vehicle mobility is random and uncorrelated with future channel state should drive the LSTM to near-chance prediction (NMSE close to 1), confirming that the result depends on the physical CAM–CSI relationship rather than on the architecture alone.","tokens_in":18904,"feed_emoji":"📡","tokens_out":6709,"duration_ms":66564,"temperature":0.7,"pith_summary":"The paper aims to show that a 5G base station can build an accurate channel-state predictor for fast-moving vehicles without any extra feedback from the vehicles, using only the cooperative awareness messages (CAMs) that vehicles already broadcast to each other at 5.9 GHz. The gNB overhears position, speed, and acceleration from nearby cars, pairs these with the channel state it measures from normal reference signals, and uses the paired data to train a lightweight LSTM that predicts the next channel state 100 ms ahead. The authors evaluate this in a realistic simulated urban mmWave scenario and report normalized mean squared error around 0.01, with gradual degradation when the prediction horizon is extended to 1 s or when noise is added to the CSI labels. If true, this gives network operators a low-overhead way to prepare beam switches proactively, and the same model can be trained across many base stations with federated learning to produce a network-wide predictor.","feed_headline":"Vehicle broadcasts alone predict 5G mmWave channel state","feed_subtitle":"A lightweight LSTM turns overheard C-V2X mobility data into proactive beam switching, no extra feedback.","key_machinery":"The load-bearing object is the CAM-to-CSI training pair: the gNB overhears Cooperative Awareness Messages (periodic 100 ms vehicle-status broadcasts over the 5.9 GHz sidelink), extracts five mobility features (x, y, z, speed, acceleration), and labels each snapshot with the concurrently measured mmWave CSI vector. This pairing lets the gNB build and refresh its own labeled dataset without any UE involvement, which in turn makes the whole prediction loop autonomous and self-trainable. The predictor itself is a single LSTM layer with 10 hidden units plus a fully connected output head producing 32 real values (the real/imaginary parts of 16 complex CSI coefficients); the LSTM carries temporal s","core_discovery":"The paper claims that the position, speed, and acceleration found in C-V2X cooperative awareness messages are sufficient inputs for a lightweight LSTM to predict future mmWave CSI accurately enough for proactive beam management. Time-aligned CAM-CSI pairs (CSI split into real and imaginary parts and normalized) form the training set. A ten-hidden-unit LSTM over ten snapshots reaches NMSE 0.0094–0.018 across four base stations, and a federated version reaches 0.0165 on a centralized test set. The predictor also tolerates CSI noise above roughly 10 dB, stays below 0.012 NMSE at a one-second horizon on BS1, and scales to eight-antenna vehicles with little accuracy loss.","pith_inferences":["The paper explicitly leaves CAM delivery delay and loss out of scope; a realistic deployment where some vehicles do not broadcast CAMs or where messages arrive late would violate the time-alignment assumption, so an end-to-end test with a real sidelink protocol (or even a simple packet-loss model) would be the natural next experiment.","Because the inputs are generic mobility features, the same framework should transfer to any source of timely vehicle state — e.g., onboard GPS reports via the network or roadside sensors — as long as the 100 ms cadence and the gNB's time-stamping are preserved.","The reported NMSE numbers are on normalized complex CSI; an unstated but testable consequence is whether a beam-index or beam-angle prediction task would inherit the same accuracy, since beam decisions depend on quantized angles rather than raw complex channel vectors.","The FL result suggests a path to a continuous self-updating network model: as CAM and CSI data accumulate, the O-RAN MLOps loop could retrain on a schedule, effectively turning the radio network into a self-tuning predictor that adapts to new street layouts or traffic patterns without manual labeling."],"forward_implications":["Proactive beam switching becomes possible with no extra uplink feedback: the gNB needs only the CAM stream it already overhears, so beam management overhead and reaction time can both be reduced.","A single lightweight model (about 2,100 parameters, ~12 microseconds inference) fits inside the near-real-time RIC constraints of O-RAN as an xApp, making deployment practical at the edge.","Federated learning can produce a global CSI predictor across multiple gNBs (NMSE 0.0165) without sharing raw CSI or mobility data, which helps when individual sites have small or non-representative datasets.","The predictor remains useful beyond the 100 ms baseline: NMSE stays below 0.012 for BS1 even at a 1 s prediction horizon, so longer-horizon resource management and handover planning become feasible.","The framework generalizes to multi-antenna vehicles: scaling from 1 to 8 receive antennas raises NMSE by only ~0.0004 on BS1 while inference time stays under 14 microseconds."],"fun_headline_variants":["C-V2X broadcasts predict mmWave CSI for proactive beam switching","Vehicle CAMs train LSTM to forecast 5G mmWave channel state","Proactive beam switching via C-V2X-driven CSI prediction in O-RAN","Federated CSI prediction from vehicle data reaches 0.016 NMSE in 5G","Self-trainable CSI prediction uses C-V2X data for O-RAN mmWave"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The framework collapses if the gNB cannot reliably overhear complete, time-synchronized CAM messages from every vehicle in its cell, since both training labels and inference inputs depend on that stream; the paper assumes this delivery and does not model message loss, delay, or non-broadcasting vehicles.","fun_headline_variants_meta":{"raw":{"variants":["C-V2X broadcasts predict mmWave CSI for proactive beam switching","Vehicle CAMs train LSTM to forecast 5G mmWave channel state","Proactive beam switching via C-V2X-driven CSI prediction in O-RAN","Federated CSI prediction from vehicle data reaches 0.016 NMSE in 5G","Self-trainable CSI prediction uses C-V2X data for O-RAN mmWave"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000942,"raw_usage":{"total_tokens":3859,"prompt_tokens":740,"completion_tokens":3119,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":3012}},"tokens_in":484,"tokens_out":3119,"duration_ms":20360,"temperature":1.0,"reasoning_tokens":3012,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T06:11:54.997199+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same urban scenario and run the predictor with a percentage of CAM messages randomly dropped or delayed by more than 100 ms; if NMSE degrades sharply (e.g., doubles) under realistic loss rates, the autonomy claim fails in practice. More directly, a synthetic dataset in which vehicle mobility is random and uncorrelated with future channel state should drive the LSTM to near-chance prediction (NMSE close to 1), confirming that the result depends on the physical CAM–CSI relationship rather than on the architecture alone.","supporting_citations":[],"review_version":1}