{"id":"051429f5-ba32-405c-87b9-1ac3a6a35491","arxiv_id":"2501.14970","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A comprehensive review of AI-driven cellular positioning organized by the 3GPP taxonomy, plus a small benchmark of four neural network models on the MaMIMO and DeepMIMO datasets.","lead":"This paper surveys how artificial intelligence is used to locate phones and devices through cellular networks, covering the basic math, the 3GPP standards that define these features, and many recent algorithms. It also runs a small benchmark comparing four neural network models on two public datasets, with code shared on GitHub, but the measured ranking contains a likely typo.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Case study's random split likely leaks spatial information, so the reported model ranking may reflect interpolation rather than generalization.","rationale":"The reader's weakest_assumption correctly flags the random split and the correctness of the reported numbers, and the '1.7 mm' typo is a real internal inconsistency. My stress-test sharpens this into a specific mechanism: spatial leakage from random splitting of densely sampled grids. This is more load-bearing than the typo alone because it threatens the validity of the paper's only quantitative evaluation and the model-ranking conclusions drawn from it. The survey's descriptive content (3GPP taxonomy, standards evolution, SOTA organization) is not invalidated by this issue, so the verdict should remain CONDITIONAL: the paper is usable as a reference once the case study is either fixed to use spatially separated splits or downgraded to an illustrative comparison without strong ranking claims. I partially agree with the reader because the spatial-leakage mechanism is not explicitly identified as the core problem, though the reader's emphasis on the single random split is consistent with it.","tokens_in":46969,"tokens_out":6498,"duration_ms":63696,"concrete_test":"Re-run the released GitHub code on both datasets under two conditions: (a) the original random split and (b) a spatially disjoint split (e.g., train on one half of the MaMIMO grid and the DeepMIMO street segments, test on the other), using at least 5 random seeds per condition and reporting median errors with interquartile ranges. If the CNN-vs-Transformer ordering is not preserved under spatial-block splits, or if the interquartile ranges overlap substantially, the case study's ranking claim and the Sec. VII.D conclusions fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's only original quantitative contribution is the Sec. VII.B case study, which reports a consistent model ranking (CNN > Transformer > LSTM > FCNN) on MaMIMO ULA-LOS and DeepMIMO O1, and Sec. VII.D uses this to conclude that CNNs perform consistently well and Transformers need more data. The evaluation, however, uses a single random split of 7000/1000/2000 samples drawn from densely and regularly spaced measurement grids: MaMIMO has sub-millimeter CNC-controlled positions over a 1.25 m x 1.25 m grid, and DeepMIMO O1 has dense street-level UE locations. Randomly assigning samples from the same or neighboring grid positions to both training and test sets creates spatial leakage, because a test CSI can be nearly identical to a training CSI at an adjacent or even identical position. This is precisely the interpolation regime where deep networks are strongest, so the reported median errors (33 mm on MaMIMO, 1.0 m on DeepMIMO) are likely optimistic, and the ranking could change under a spatially disjoint split. The paper provides no error bars, multiple seeds, or hyperparameter details, so the ranking is not established. The additional '1.7 mm' LSTM value on DeepMIMO is almost certainly a units typo, further undermining confidence in the reported numbers. Since the case study underpins the survey's only quantitative performance-evaluation claim and its model-generalization lessons, this is the most load-bearing weak point in an otherwise descriptive survey.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey of AI-driven cellular positioning. It reviews the fundamentals of wireless positioning and of mainstream AI models, traces the evolution of 3GPP positioning standards with emphasis on Release 18 AI/ML positioning and the five deployment cases, and organizes the state of the art into AI/ML-assisted positioning (LOS/NLOS detection, TOA/TDOA estimation, angle estimation) and direct AI/ML positioning (fingerprinting, knowledge-assisted learning, channel charting). It also compares public datasets and presents a case study on MaMIMO and DeepMIMO that benchmarks FCNN, CNN, LSTM, and Transformer models, with the stated finding that CNNs perform most consistently while Transformers require more data. The paper's central claim is that it provides a comprehensive, up-to-date, 3GPP-structured reference map of the field.","tokens_in":47249,"tokens_out":2047,"duration_ms":21237,"significance":"If the survey's descriptive content is accurate, it fills a useful niche: it is one of the few surveys that couples the 3GPP AI/ML positioning taxonomy with a systematic review of both assisted and direct AI/ML methods, and it releases open-source evaluation code for its case study. The dataset comparison table and the summary of 3GPP deployment cases (Case 1--3b) are potentially valuable reference material for researchers entering the area. However, the paper's only original quantitative contribution, the case study in Sec. VII.B, has methodological flaws that directly affect the model-ranking claims drawn in Secs. VII.B and VII.D. The descriptive survey parts are broadly sound, but the empirical section needs substantial revision before the manuscript can be accepted.","major_comments":[{"comment":"The case study uses a single random 7000/1000/2000 split on spatially dense measurement grids: MaMIMO positions lie on a 1.25 m x 1.25 m grid with sub-millimeter CNC-controlled accuracy, and DeepMIMO O1 UEs are densely spaced along street-level grid points. A random split assigns samples from identical or adjacent grid positions to both training and test sets, so the test CSI can be nearly identical to a training CSI. This spatial leakage makes the reported median errors (33.1 mm on MaMIMO, 1.0 m on DeepMIMO for CNN) optimistic and potentially changes the ranking under a spatially disjoint split. Please report results with a spatially disjoint split (e.g., train and test on separated grid regions or with a minimum position separation), multiple seeds, and confidence intervals or per-seed spreads.","section":"Sec. VII.B"},{"comment":"The LSTM median error on DeepMIMO is reported as 1.7 mm, which is physically implausible for the configuration used (20 MHz bandwidth, 100 subcarriers, 64-element ULA in an urban street scenario) and contradicts both the CDF in Fig. 11 and the other models' errors in the same setting. This is almost certainly a units typo (1.7 m), but the presence of such an error in the paper's only quantitative table undermines confidence in the remaining reported numbers. Please correct the value and audit the other entries for consistency with the figures.","section":"Sec. VII.B, Fig. 11"},{"comment":"The general conclusions that 'CNNs perform consistently well' and 'Transformers require more data' are supported only by a single configuration (fixed 20 MHz, 100 subcarriers, 64-element ULA, one data split, one scenario per dataset) with no hyperparameter details, no model-size or training-budget matching, and no statistical significance testing. These claims are therefore not established by the presented evidence. Please either soften the conclusions to be explicitly configuration-specific or add the missing experimental controls and uncertainty quantification.","section":"Secs. VII.B and VII.D"}],"minor_comments":[{"comment":"The text states that xG-Loc is 'an open dataset' and 'open source', but Table XV lists xG-Loc's 'Open Source Type' as 'No'. Please reconcile this inconsistency.","section":"Sec. VII.A, Table XV"},{"comment":"The figure and the associated claim that more than half of wireless positioning papers use AI by 2023 are based on Google Scholar data, but no search strings, inclusion criteria, or validation methodology are described. Adding a brief methodology note would improve reproducibility.","section":"Sec. I, Fig. 1"},{"comment":"The Transformer complexity expression uses the notation d_LSTM for the feature dimension; this should be d_Trans or another symbol to avoid confusion with LSTM dimensions.","section":"Sec. III.A.4"},{"comment":"In the text and table, 'V AE' appears with an extra space; also 'angle-delay cannel power matrix' should be 'channel power matrix' (ADCPM). Please fix these typos.","section":"Sec. V.B, Table VIII"},{"comment":"There are several spacing and formatting issues (e.g., 'UA Vs', 'ttoa' in the TOA equation, and 'V oronoi' in Sec. VI.A.2) that should be corrected in a final pass. The reference list also contains some garbled author names (e.g., 'V . D' in [95]) that need cleanup.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The survey portion of the paper is a useful reference and the 3GPP taxonomy is well chosen. The main blocker is the case study: its spatial-leakage-prone evaluation design and the obvious '1.7 mm' unit error mean the quantitative claims cannot be trusted as-is. The fixes are relatively local (re-run with disjoint splits, correct the typo, add uncertainty analysis), which is why I recommend major revision rather than rejection. I would also gently encourage the editors to ensure the xG-Loc open-source inconsistency is resolved, as it is visible to the community."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, standard survey of AI-driven cellular positioning, organized around the 3GPP R18 taxonomy. Its real value is as a reference map. The standards recap (deployment cases, KPIs, lifecycle management) is accurate and useful, and the SOTA tables give a newcomer a fast route into the literature. The dataset comparison table is also handy, and the authors actually shipped code for the small benchmark, which is more than most surveys do.\n\nThe genuinely new artifact is the Sec. VII.B case study comparing FCNN, CNN, LSTM, and Transformer on MaMIMO and DeepMIMO. That is where the soft spots are. The random 7000/1000/2000 split on dense, regularly spaced grids is a real problem. MaMIMO positions are CNC-controlled with sub-millimeter precision over a 1.25 m square; DeepMIMO O1 is a dense street-level grid. Randomly assigning samples from the same or neighboring grid positions to train and test means the test set measures interpolation, not generalization. The reported ranking (CNN > Transformer > LSTM > FCNN) is likely optimistic and could change under a spatially disjoint split. There are no error bars, no multiple seeds, and no hyperparameter detail. The LSTM median on DeepMIMO is reported as '1.7 mm' against meters for everything else, which has to be a units typo. These are fixable, but they matter because the paper uses this benchmark to draw a general lesson about model choice.\n\nTwo smaller inconsistencies: Table XV labels xG-Loc 'No' for open source while the text calls it an open dataset, and Fig. 1's Google Scholar methodology is not described, so the 'more than half of papers use AI by 2023' claim is hard to verify.\n\nNone of this destroys the survey's descriptive value. The classification into AI/ML-assisted vs. direct AI/ML positioning, the LOS/NLOS, TOA/TDOA, angle, fingerprinting, and channel charting summaries, and the 3GPP discussion are competent and mostly accurate. The citation coverage looks broad and appropriate, with no sign of promotional self-citation. It is not a groundbreaking paper, but it does not need to be.\n\nWho benefits: graduate students and engineers entering the field, and anyone needing a compact R18-oriented map. It deserves a serious referee, but a referee should insist on fixing the split protocol, adding seeds and error bars, correcting the typo, and reconciling the xG-Loc entry.","headline":"Useful 3GPP-centered survey with a credible taxonomy; its own benchmark is undermined by spatial leakage in the data split.","tokens_in":47790,"tokens_out":3313,"would_cite":true,"duration_ms":30909,"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 claims that AI-driven cellular positioning is now a coherent field best understood through the 3GPP taxonomy of AI/ML-assisted versus direct AI/ML positioning, and that a CNN baseline beats Transformer, LSTM, and FCNN on both…","keywords":["wireless positioning","AI-driven cellular positioning","3GPP AI/ML positioning standards","AI/ML-assisted positioning","direct AI/ML positioning","channel charting","fingerprint-based localization","positioning datasets"],"falsifier":"Running the released code with the stated data split and hyperparameters and checking whether the median errors reproduce 141.2 mm, 33.1 mm, 51.8 mm, and 40.9 mm on MaMIMO and 13.1 m, 1.0 m, 1.7 m (not mm), and 1.2 m on DeepMIMO; a mismatch in the LSTM DeepMIMO value, or in the CNN/Transformer ordering under multiple random seeds, would falsify the paper's ranking. Separately, recovering the Google Scholar search behind Fig. 1 would determine whether the 'more than half of positioning papers use AI by 2023' statistic is reproducible.","tokens_in":46736,"feed_emoji":"📍","tokens_out":7177,"duration_ms":60414,"temperature":0.7,"pith_summary":"Wireless positioning is shifting from geometric models to machine learning, and this paper tries to give researchers a reliable map of that shift: the fundamentals of radio positioning, the AI models applied to it, the 3GPP standards that now govern AI/ML positioning, the state of the art in both assisted and direct AI approaches, and the datasets available for training and testing. The organizing claim is that 3GPP's distinction between AI/ML-assisted positioning (AI refines measurements such as LOS/NLOS, TOA/TDOA, and angles, which feed a conventional solver) and direct AI/ML positioning (AI maps signals straight to coordinates via fingerprinting, knowledge-assisted learning, or channel charting) covers the field. The paper's own experiment compares FCNN, CNN, LSTM, and Transformer on the MaMIMO and DeepMIMO datasets and reports that CNN ranks first on both, with median errors of 33.1 mm on MaMIMO ULA-LOS and 1.0 m on DeepMIMO O1. If the survey is right, it gives practitioners a common vocabulary and a baseline comparison to build on; what the field still lacks, the paper argues, are datasets with mobility, Doppler, and multi-modal data, and a unified evaluation benchmark.","feed_headline":"CNN beats Transformer, LSTM, FCNN in positioning test","feed_subtitle":"Survey's case study puts CNN first with 33.1 mm median error on MaMIMO and 1.0 m on DeepMIMO.","key_machinery":"The central organizing object is the 3GPP AI/ML positioning taxonomy from the Release 18 study: AI/ML-assisted positioning (models output refined measurements or probabilities such as LOS/NLOS indicators, TOA/TDOA, or AOA/AOD that feed a conventional positioning solver) versus direct AI/ML positioning (models consume raw channel observations and output the UE location directly). Around this, the paper builds its review around the lifecycle management (LCM) framework—data collection, model training, management, inference, and storage—and the five deployment cases (model on UE, gNB, or LMF) that determine communication overhead and scalability. The other load-bearing mechanism is the case-study comparison of FCNN, ResNet-32 CNN, LSTM, and Transformer on MaMIMO and DeepMIMO, which is what grounds the survey's claims about model selection. The channel model formula $H_s = \\sum_{l=1}^L \\alpha_l \\mathbf{a}(\\theta_l,\\varphi_l)e^{-j2\\pi f_s \\tau_l}$ defines the CSI object that nearly every surveyed method uses as input.","core_discovery":"The central claim is that AI-driven cellular positioning has matured into an identifiable technical area with a 3GPP-defined structure, and that current research divides into two classes: AI/ML-assisted positioning, where models improve LOS/NLOS detection, TOA/TDOA estimation, or angle prediction and hand those refined measurements to a conventional geometry-based location estimator, and direct AI/ML positioning, where models learn the signal-to-position mapping end to end through fingerprinting, knowledge-assisted learning, or channel charting. The paper argues that model-based positioning struggles under multipath, NLOS, hardware impairment, and database aging, while AI models learn those patterns from data, and that 3GPP Release 18 has begun standardizing this through a lifecycle-management framework and five deployment cases spanning UE, gNB, and LMF. Its original quantitative contribution is a comparison of four backbones on one measured and one simulated dataset (MaMIMO ULA-LOS and DeepMIMO O1 at 20 MHz, 100 subcarriers, 64-element ULA): median errors on MaMIMO of 141.2 mm (FCNN), 33.1 mm (CNN), 51.8 mm (LSTM), and 40.9 mm (Transformer), and on DeepMIMO of 13.1 m, 1.0 m, 1.7 mm (written in the text, though the figure indicates 1.7 m), and 1.2 m, giving the ordering CNN, Transformer, LSTM, FCNN. The survey concludes that datasets are the bottleneck: none of the reviewed public datasets combines mobility, Doppler, multi-band, and multi-modal sensing, so it calls for standardized benchmarks and extensible simulators.","pith_inferences":["I would not bet the CNN-over-Transformer ranking on a single split: the paper does not report seed variance or hyperparameter tuning, so the ordering may shift under repeated runs; that is my inference, not the paper's claim.","The '1.7 mm' LSTM value deserves a policy response: the field needs a convention that error units are stated explicitly and that tables and CDF figures are cross-checked, because unit typos of this kind make model comparisons meaningless.","A natural next experiment the paper does not run is to pretrain a Transformer on unlabeled CSI from the same scenarios (the LWLM-style self-supervised approach the paper reviews) and then compare against ResNet CNN at the same label budget; given the paper's 'Transformers need more data' explanation, this would directly test whether the ranking persists.","If the 3GPP taxonomy is as complete as the survey claims, future standards work should define evaluation metrics that penalize communication overhead and on-device compute, not just accuracy, because Cases 2b and 3b trade overhead for accuracy."],"forward_implications":["A common vocabulary: the AI/ML-assisted versus direct AI/ML split, tied to 3GPP deployment cases, gives researchers and operators a standard way to classify positioning algorithms across papers and releases.","On the two tested datasets, CNN (ResNet-32) is the best default backbone for small-data, single-BS, static-channel positioning; Transformer needs more data or pretraining to catch up.","Simulated ray-tracing data (DeepMIMO) preserves the relative ranking of models found on real measurements (MaMIMO), but absolute errors differ by orders of magnitude, so synthetic benchmark numbers should not be treated as achievable field accuracy.","The biggest bottleneck identified is data: until datasets include user mobility, Doppler, multi-band signals, and vision/IMU modalities, AI positioning models cannot be fairly evaluated for dynamic 6G scenarios."],"supporting_citations":[{"why":"Survey of cellular localization from 1G to 5G that the standards-evolution section builds on.","marker":"[1]"},{"why":"Earlier machine-learning-based localization survey that the paper positions itself against by adding the 3GPP taxonomy and a cellular-dataset focus.","marker":"[15]"},{"why":"Tutorial on 5G positioning that supplies background for ranging, angle, and standards techniques.","marker":"[21]"},{"why":"3GPP specification defining 5G positioning service levels and KPIs that ground the requirements tables.","marker":"[195]"},{"why":"3GPP Release 18 AI/ML positioning study that supplies the direct-versus-assisted taxonomy and the UE/gNB/LMF deployment cases.","marker":"[200]"},{"why":"MaMIMO measured indoor CSI dataset used for the ULA-LOS case study.","marker":"[349]"},{"why":"DeepMIMO ray-tracing simulator and dataset used for the O1 case study.","marker":"[350]"},{"why":"Channel charting overview that anchors the channel-charting subsection and its notation.","marker":"[111]"},{"why":"Large wireless localization model study that the survey draws on for self-supervised pretraining and foundation-model opportunities.","marker":"[201]"}],"fun_headline_variants":["Survey: AI positioning hinges on datasets, not models","CNN tops Transformer, LSTM in 3GPP positioning survey","AI-driven positioning: 3GPP standard, two classes, one bottleneck","Positioning survey: CNN leads, but datasets lag behind","Wireless positioning meets AI: survey maps standards and gaps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Sec VII.B case study—four architectures, one measured and one simulated dataset, a single random 7000/1000/2000 split, and fixed bandwidth and array settings—is a fair and representative basis for the paper's model ranking, and that the reported median errors are accurate, an assumption the DeepMIMO LSTM '1.7 mm' value already violates inside the paper.","fun_headline_variants_meta":{"raw":{"variants":["Survey: AI positioning hinges on datasets, not models","CNN tops Transformer, LSTM in 3GPP positioning survey","AI-driven positioning: 3GPP standard, two classes, one bottleneck","Positioning survey: CNN leads, but datasets lag behind","Wireless positioning meets AI: survey maps standards and gaps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000715,"raw_usage":{"total_tokens":3332,"prompt_tokens":1179,"completion_tokens":2153,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":795,"completion_tokens_details":{"reasoning_tokens":2067}},"tokens_in":795,"tokens_out":2153,"duration_ms":14410,"temperature":1.0,"reasoning_tokens":2067,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:44:37.224388+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Running the released code with the stated data split and hyperparameters and checking whether the median errors reproduce 141.2 mm, 33.1 mm, 51.8 mm, and 40.9 mm on MaMIMO and 13.1 m, 1.0 m, 1.7 m (not mm), and 1.2 m on DeepMIMO; a mismatch in the LSTM DeepMIMO value, or in the CNN/Transformer ordering under multiple random seeds, would falsify the paper's ranking. Separately, recovering the Google Scholar search behind Fig. 1 would determine whether the 'more than half of positioning papers use AI by 2023' statistic is reproducible.","supporting_citations":[],"review_version":1}