{"id":"0d4ebd82-7577-4b44-a94a-a06a91f08bdd","arxiv_id":"2411.17716","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A U-Net trained on other access points' channel gain maps can infer a new access point's channel knowledge map from its location and the surrounding maps.","lead":"This paper trains a U-Net to create a channel knowledge map for a new access point using the channel maps of nearby existing access points plus the new point's coordinates. If it works in practice, dense networks could skip costly per-point measurements.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation never tests a genuinely new AP location: both training and validation use the 80 pre-existing AP positions as targets, so the reported 2.38 dB does not yet establish cross-AP inference for arbitrary AP placements.","rationale":"The reader's weakest_assumption concerned generalization from simulated RadioMapSeer environments to real deployments. That is a valid external-validity concern, but I find a more immediate and more easily testable gap: the evaluation never exercises the actual inference scenario of a new AP at a location not among the 80 predefined AP positions. Since both training and validation draw targets only from those positions, the reported accuracy is an interpolation result over the training target-location distribution, not evidence for arbitrary c0. This gap is internal to the paper's own setup and can be settled with a held-out-position experiment on the same dataset, without needing measured data. I do not think this disproves the method; the U-Net may well generalize to unseen positions because the target location is provided as an input channel. But the paper currently provides no evidence for that, and the central claim is phrased in terms of potentially new APs. The paper otherwise has merit: it uses a public dataset, the method is simple, and the qualitative figures suggest the model captures building-occlusion features. The missing ablations, unreported blend weight omega, and lack of code are secondary to this validation gap. I therefore keep the reader's conditional verdict unchanged, with the added condition that unseen-AP-location generalization be demonstrated.","tokens_in":7438,"tokens_out":7048,"duration_ms":73478,"concrete_test":"Hold out AP positions rather than entire environments: for each RadioMapSeer map, train the UNet using, say, 60 of the 80 APs as targets and reserve the remaining 20 AP positions as unseen targets. At test time, feed the held-out target's location map plus the CKMs of the other 79 APs and compute the RMSE on the 20 held-out CGMs. Repeat over multiple random splits and report mean and error bars. If the held-out-position RMSE is close to the reported 2.38 dB, the concern is resolved; if it is substantially higher, the central claim must be restricted to AP positions that resemble training locations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central use case is generating a CKM for a potentially new AP at an arbitrary location c0 from the CKMs of existing APs, as formulated in Eq. (2). However, Algorithm 1 and the validation protocol in Section IV.A always choose the target AP from the 80 AP positions already defined in each RadioMapSeer environment. The target location map M*_AP,target is therefore always one of the discrete positions used to generate the dataset; no held-out AP location is ever evaluated. The 2.38 dB RMSE in Table I is thus measured only for target positions drawn from the same distribution as the training targets, not for genuinely new placements. The model could succeed by exploiting the strong similarity between the target CGM and CGMs of nearby existing APs, or by memorizing the set of target coordinate patterns, without learning a generalizable cross-AP mapping. This is an internal-validity gap that exists even before considering transfer from simulated RadioMapSeer environments to measured data. A related consequence is that the architecture's fixed 80 input channels also means variable numbers of existing APs, which a dense network would have, are not tested. The central claim of feasibility for arbitrary new APs therefore rests on an untested extrapolation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the problem of generating a channel knowledge map (CKM) for a potentially new access point (AP) in a dense network, using only the CKMs of existing APs and the new AP's location. The authors propose a UNet-based deep learning method: the input is a stack of feature maps formed by combining each existing AP's channel gain map (CGM) with a one-hot AP location map, plus a pre-convolved target AP location map; the output is the target AP's CGM. The model is trained and evaluated on the RadioMapSeer dataset (500 simulated environments for training, 100 for validation, 80 APs per environment, each AP in turn as target). The reported RMSE is 2.38 dB, which the authors compare to RadioUNet's 2.03 dB and to two simpler baselines (weighted CKM inference: 5.30 dB; 3GPP path-loss model: 35.72 dB). The central claim is that cross-AP CKM inference is feasible and effective without requiring the physical environment map.","tokens_in":7684,"tokens_out":6662,"duration_ms":66274,"significance":"If the reported result holds, the proposed method could reduce the overhead of CKM construction and update in dense networks such as cell-free systems, and the problem formulation (cross-AP inference from other APs' CKMs) is a useful extension of existing CKM literature. The paper contributes a concrete UNet architecture, a training procedure on a public dataset, and a comparison against baselines. It also explicitly avoids using the physical environment map as input, which is a meaningful difference from RadioUNet. However, the significance is conditional on addressing the evaluation gaps described below; as it stands, the central feasibility claim is not yet convincingly established for the stated problem of arbitrary new AP locations. The work is an incremental but potentially useful contribution if the validation is strengthened.","major_comments":[{"comment":"The evaluation protocol never tests a genuinely new AP location: both training and validation always choose the target AP from the 80 pre-existing AP positions defined in each RadioMapSeer environment. The problem formulation in Eq. (2) explicitly considers a potentially new AP at an arbitrary location c0, but the reported 2.38 dB RMSE in Table I is measured only for target positions drawn from the same discrete set used to generate the dataset. The model could succeed by exploiting the strong similarity between the target CGM and CGMs of nearby existing APs, or by memorizing target coordinate patterns, without learning a generalizable cross-AP mapping. A proper evaluation should hold out a set of AP coordinates (or generate new transmitter positions) and test the trained model on those locations in unseen environments.","section":"Section IV.A, Algorithm 1"},{"comment":"The blend weight ω in Eq. (4) is a free parameter that controls the relative contribution of the AP location map and the channel gain map in the input feature representation, but its value is never reported anywhere in the paper. Since the input to the network depends directly on ω, the experimental results are not reproducible without this value. Please report the chosen ω and, ideally, an ablation showing the sensitivity of the results to this parameter.","section":"Section III.A, Eq. (4)"},{"comment":"The model is selected on the validation set (\"the model that minimized the MSE loss in the validation set over the 15 epochs is saved\"), and the same validation set is then used to produce the results in Table I. This selection procedure can bias the reported performance upward because the validation set has been used for model choice. A separate test set, or a nested validation approach, should be used to report the final numbers.","section":"Section IV.A"},{"comment":"The results are from a single training run with no error bars, seed variation, or confidence intervals. Deep-learning training is stochastic, and the reported improvement of the proposed method (2.38 dB) over the weighted baseline (5.30 dB) could be affected by training variance. Please report the mean and standard deviation over multiple runs, or at least a bootstrap over environments.","section":"Section IV.B, Table I"},{"comment":"No ablation is provided to separate the contribution of the target AP location channel from the contribution of the other APs' CGMs in the input. The input contains both the pre-convolved target location map and the 80 other-AP feature maps, so it is unclear whether the model is actually exploiting the cross-AP correlation or simply learning a location-to-CGM mapping from the target location channel alone (or from a trivial average of the other CGMs). A baseline that uses only the target location (e.g., predicting the mean CGM over the training set) or only the other APs' CGMs without the target location would clarify what the model learns and would directly test the paper's core hypothesis.","section":"Section III.B, Section IV"}],"minor_comments":[{"comment":"The notation \"M_input ∈ R^{W^2×(n+1)}\" uses a lowercase n that is not defined; it should be N, the number of existing APs.","section":"Section III.A"},{"comment":"There is a typo: \"c is any poss coordinate in the CKM\" should read \"c is any possible coordinate in the CKM.\"","section":"Section III.A"},{"comment":"The comparison with RadioUNet is not a like-for-like baseline: RadioUNet takes the physical environment map as input and addresses a different task (radio map estimation from the environment and transmitter location). The statement that the proposed method is \"on the same level\" as RadioUNet should be qualified to avoid implying a direct algorithmic comparison.","section":"Section IV.B, Table I"},{"comment":"The 3GPP TR 38.901 path-loss baseline is not fully specified (carrier frequency, antenna heights, LOS/NLOS assumptions, etc.), which makes the reported 35.72 dB RMSE difficult to interpret or reproduce.","section":"Section IV.B, Benchmark 2"},{"comment":"The paper notes that for other datasets the number of input channels must be adjusted, but it does not discuss how the method scales to a variable number of existing APs, which is a natural requirement in dense networks. A brief discussion or a small experiment with a different N would strengthen the generality claim.","section":"Section III.B"}],"recommendation":"major_revision","confidential_remarks":"The paper comes from a well-known group in the CKM area and uses a public dataset, which is a plus. The core idea is reasonable, but the evaluation does not currently match the stated problem: the central claim of generating CKMs for arbitrary new AP locations is not tested because only the 80 predefined AP positions serve as targets. The unreported ω and the validation-based model selection are also reproducibility concerns. I believe these issues are fixable within the scope of the paper, so I recommend major revision rather than rejection. If the authors can add an evaluation on held-out AP locations, report ω, and provide error bars or a separate test set, the paper would be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper does something genuinely new: it generates the full channel gain map for a target AP from the maps of other APs plus the target location, using a U-Net. Previous cross-AP work inferred channel state at a single location, not a whole map. Second, the validation has a blind spot that matters: the test always chooses target APs from the same 80 pre-existing AP positions used to build the dataset, so a truly new AP placement is never tried. The 2.38 dB number is real for the tested setting, but the paper's stated use case—arbitrary new APs—is an extrapolation.\n\nWhat's solid: the task is well motivated for cell-free networks, the formulation in Eq. (2) is clean, and the benchmarks are appropriate. Beating distance-weighted inference by ~3 dB and a 3GPP path-loss model by ~33 dB is a real result. Using 500/700 RadioMapSeer environments for training and 100 disjoint ones for validation is a reasonable split. The U-Net design is standard, but the pre-convolution for the sparse target location map is a sensible tweak.\n\nSoft spots, in order of severity. The biggest is the held-out-location gap; it should be easy to fix: hold out some AP positions entirely (say, 10 of the 80) and only test on those. Second, omega in Eq. (4) is not reported anywhere; the input representation literally depends on a chosen weight and the reader cannot reproduce the numbers without it. Third, there is a single training run and no error bars; with a deep model and 40k training samples, run-to-run variance could be a couple tenths of a dB. Fourth, no ablation separates what the target location map contributes versus the other APs' CGMs; the model could be leaning heavily on the location prior. The comparison to RadioUNet is fair only as a rough benchmark since RadioUNet solves a different task (radio map from environment), not the same cross-AP problem. Minor: no code.\n\nWho it's for: people building CKM construction pipelines and anyone interested in environment-aware 6G. The central idea is plausible and worth engaging with, but the paper as written is a feasibility study that needs a tighter validation before the cross-AP generation claim should be taken at face value.\n\nRecommendation: yes, send to review, but expect the reviewers to ask for the held-out-location test, the omega value, and error bars. It's a useful advance, not a breakthrough.","headline":"Genuinely new task (full cross-AP CKM generation) with a real result on RadioMapSeer, but the validation never tests a truly new AP location, so the central deployment claim rests on an untested extrapolation.","tokens_in":8181,"tokens_out":2062,"would_cite":false,"duration_ms":20278,"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 aims to show that a deep network can generate a new access point's channel knowledge map purely from the channel knowledge maps of neighboring access points and the new AP's location, without any physical environment map or new…","keywords":["channel knowledge map","cross-AP CKM inference","UNet","deep learning","cell-free networks","channel gain map","RadioMapSeer","environment-aware communication"],"falsifier":"Take a real or ray-traced dense deployment, construct CKMs for a handful of APs by measurement or simulation, infer the CKM of a held-out AP with the trained UNet, and compare against the measured CKM. If the RMSE on data outside RadioMapSeer-style synthetic cities is substantially worse than 2.38 dB, or no better than the distance-weighted baseline, the central claim of learnable, transferable cross-AP correlation is falsified.","tokens_in":7256,"feed_emoji":"📡","tokens_out":11116,"duration_ms":88593,"temperature":0.7,"pith_summary":"The paper tries to establish that a channel knowledge map (CKM), a location-specific database of channel properties, for a newly introduced access point can be generated from the CKMs of existing access points in the same region plus the new AP's location, without measuring or ray-tracing the new AP. The enabling premise is that APs sharing one physical environment have correlated, location-dependent radio behavior, because buildings and terrain shape every AP's signal field in related ways. If this works, dense networks such as cell-free systems can build CKMs for many APs from measurements at only a few, update CKMs cheaply, and evaluate candidate AP locations without site surveys. The paper reports that a UNet trained on simulated city environments infers the target AP's CKM with an RMSE of 2.38 dB, on par with the map-based RadioUNet benchmark (2.03 dB) even though the proposed method never sees the physical environment map.","feed_headline":"AI predicts a new base station's signal map from neighbors' maps","feed_subtitle":"No map of the physical environment is needed; error is 2.38 dB, nearly matching map-based RadioUNet.","key_machinery":"The load-bearing machinery is a UNet whose input has $N+1$ channels: one channel is the target AP's location map, pre-convolved with a $3 \\times 3$ all-ones kernel to widen the sparse location peak, and the other $N$ channels are per-AP feature maps formed by a weighted sum of each existing AP's channel gain map and its one-hot location map, $M_n = (1-\\omega)G_n + \\omega \\mathcal{M}_{\\mathrm{AP},n}$. During training each AP in turn acts as the target and the remaining APs become inputs, with mean squared error between the inferred and ground-truth channel gain maps as the loss. The network uses $5 \\times 5$ convolutions to capture building-edge features and dimensionality-reduction convolutions to handle the 80-channel input efficiently. At inference, the new AP's location and the existing APs' CKMs produce the complete target CKM, with no environment map and no onsite measurements.","core_discovery":"The central discovery is that the implicit cross-AP correlation in channel knowledge can be learned end-to-end by a supervised UNet, and the learned mapping transfers to held-out simulated environments. On 100 validation environments disjoint from the 500 training environments, the inferred CKMs achieve 2.38 dB RMSE. Compared with a distance-weighted average of the other APs' CKMs (5.30 dB) and with the 3GPP TR 38.901 urban-microcell path-loss model (35.72 dB), the proposed inference is about 3 dB and 33 dB more accurate, respectively. The inferred maps preserve the target AP's location peak and reproduce attenuation and sharp shadowing at building edges, despite never receiving the environment map as input.","pith_inferences":["If the learnable cross-AP correlation is as strong as the simulated results suggest, the same input structure could transfer CKMs across carrier frequencies, because the shared physical layout should imprint similar structure on channel maps at different bands; this is not tested in the paper.","The multi-AP CKM input behaves as an implicit, learned surrogate for the physical environment, so the UNet's internal features could plausibly support other environment-aware tasks such as localization or coverage planning.","The likely failure mode is spatial decorrelation: with widely spaced APs or heterogeneous terrain, the CKMs may share too little usable information; a stress test with sparse AP deployment would show whether accuracy degrades toward the weighted-average baseline.","Because RadioMapSeer supplies simulated data only, the 2.38 dB accuracy should be read as a simulation-level result; a measurement campaign comparing inferred and measured CKMs in an urban microcell would settle the real-world claim."],"forward_implications":["A dense network can build CKMs for only a subset of APs by measurement or ray tracing and synthesize CKMs for the rest, cutting the initial CKM construction overhead.","Candidate locations for new APs can be scored by their inferred CKMs before deployment, enabling environment-aware AP placement in cell-free networks.","Adding or moving an AP no longer requires remeasuring the whole region: the new CKM can be inferred from the unchanged neighboring APs, making updates cost-effective.","The accuracy gap between map-free and map-based CKM generation nearly closes, so environment-aware communication becomes feasible where physical environment maps are unavailable or outdated.","With channel-count adjustments, the same input construction and UNet design apply to other CKM datasets (e.g., CKMImageNet) and to other channel quantities stored in CKMs, such as time of arrival or angle of arrival."],"supporting_citations":[{"why":"Defines the channel knowledge map concept and motivates location-specific prior channel knowledge for environment-aware communication.","marker":"[1]"},{"why":"Supplies the premise that the wireless environment is an outward manifestation of the shared physical environment, which underpins the cross-AP correlation.","marker":"[2]"},{"why":"RadioUNet provides the map-based CKM generation benchmark (2.03 dB RMSE) that the proposed method is compared against.","marker":"[5]"},{"why":"Provides prior evidence of CSI dependence across APs via mutual information, supporting the feasibility of cross-AP inference.","marker":"[7]"},{"why":"Supplies the UNet architecture that the cross-AP inference network is built on.","marker":"[9]"},{"why":"RadioMapSeer is the dataset of 700 simulated environments with 80 AP channel gain maps each, used for training and validation.","marker":"[10]"},{"why":"3GPP TR 38.901 urban-microcell path-loss model serves as the model-based benchmark that the proposed inference outperforms by about 33 dB.","marker":"[12]"}],"fun_headline_variants":["AI predicts a new AP's radio map using neighbors' data","Cross-AP deep learning predicts signal maps for new APs","No environment map needed: AI generates radio maps from neighbors","AI infers new AP signal maps from existing APs' data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the implicit correlation between CKMs of different APs in the same physical environment is strong and stable enough to be learned from 500 simulated environments and then to hold for real, unseen deployments, a transfer that the paper does not quantify with measured data.","fun_headline_variants_meta":{"raw":{"variants":["AI predicts a new AP's radio map using neighbors' data","Cross-AP deep learning predicts signal maps for new APs","No environment map needed: AI generates radio maps from neighbors","AI infers new AP signal maps from existing APs' data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00081,"raw_usage":{"total_tokens":3572,"prompt_tokens":979,"completion_tokens":2593,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":2522}},"tokens_in":595,"tokens_out":2593,"duration_ms":65239,"temperature":1.0,"reasoning_tokens":2522,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:55:20.586517+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real or ray-traced dense deployment, construct CKMs for a handful of APs by measurement or simulation, infer the CKM of a held-out AP with the trained UNet, and compare against the measured CKM. If the RMSE on data outside RadioMapSeer-style synthetic cities is substantially worse than 2.38 dB, or no better than the distance-weighted baseline, the central claim of learnable, transferable cross-AP correlation is falsified.","supporting_citations":[{"cited_title":"U-Net: Convolutional net- works for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the UNet architecture that the cross-AP inference network is built on."},{"cited_title":"Study on channel model for frequencies from 0.5 to 100 GHz,","cited_arxiv_id":null,"evidence_quote":"3GPP TR 38.901 urban-microcell path-loss model serves as the model-based benchmark that the proposed inference outperforms by about 33 dB."}],"review_version":1}