{"id":"1c666e4c-f7f2-4916-9129-c6e3a993b214","arxiv_id":"2501.06604","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A conditional denoising diffusion model generates radio maps from sparse RSS fragments or transmitter coordinates, outperforming cGAN and pix2pix in synthetic indoor and outdoor scenarios.","lead":"This paper proposes RM-Gen, a diffusion model that creates full wireless coverage maps from a small number of signal strength measurements or from transmitter locations. It reports higher accuracy than two GAN-based baselines on indoor 60 GHz and outdoor sub-6 GHz simulated scenarios.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The RSS-fragment condition encoder in Sec. IV.B omits fragment coordinates, so the model as described cannot know where each measured patch belongs; the reported 96.85% accuracy at 5% fragments is unexplained without a hidden location signal.","rationale":"The reader's weakest_assumption focuses on synthetic ray-tracing ground truth and the unspecified train/test split. Both are real reporting gaps, but they attack the external validity of the evaluation. The more load-bearing issue is internal to the method: the conditional encoder for partial RSS fragments, as written, discards the spatial locations of the fragments. Since the central claim is that RM-Gen reconstructs a specific radio map from sparse measured fragments, the model must know where each fragment was measured. The paper's Section IV.B gives no mechanism for this, and Algorithm 3's subarea selection is not connected to the encoder input. This makes the headline accuracy numbers unexplainable under the described architecture, independent of whether the ray-tracing data are realistic. I therefore disagree with the reader's identification of the weakest assumption, while agreeing with the conditional verdict: the paper needs either a corrected description of the condition encoder or a clarification of how fragment locations are provided. The proposed permutation test would settle whether the reported performance is genuine and whether the missing location information is a presentation error or a fundamental flaw.","tokens_in":18474,"tokens_out":7314,"duration_ms":77861,"concrete_test":"Obtain the authors' implementation or ask them to specify the exact input tensor to the RSS-fragment encoder e_theta(c). Then run a position-permutation test on the RM-In test set at 10% fragments: take a fixed set of fragment RSS values and assign them to (a) their true locations and (b) randomly permuted locations, keeping everything else identical. A position-aware model's accuracy should collapse in (b); the value-only encoder in Eqs. (15)-(16) would be invariant to the permutation and cannot explain the reported ~97% accuracy. If (b) also yields high accuracy, the task or evaluation is degenerate; if (a) works and (b) fails, the paper must document the location channel. This single check distinguishes a missing description from a broken method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central capability claim depends on conditioning on sparse RSS fragments, but the encoder described in Sec. IV.B cannot use their positions. Equations (15)-(16) flatten each fragment and concatenate the resulting vectors; the condition vector V therefore contains only RSS values. No fragment coordinate, subarea index, or spatial mask is fed to e_theta(c). The environment-aware selector (Algorithm 3) chooses fragments centered at subareas S_j, but those S_j are used only to pick the values and are not included in the condition. A model conditioned on V alone is permutation-invariant across fragments and has no way to align a measured -80 dBm patch to its true location in the N by N map. Without that alignment, reproducing a specific ground-truth map from 5% fragments at 96.85% accuracy (Table II) is not credible under the described architecture. Either the implementation silently feeds positions (e.g., as coordinate channels or a masked map) and the paper omits this essential design, or the evaluation protocol fixes fragment locations so that location is inferable from the training distribution. In both cases the central claim is not supported by the text as written. This is more load-bearing than the ray-tracing realism concern: even if Wireless InSite maps are perfect, this under-specification blocks the claimed mechanism.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"RM-Gen is a conditional denoising diffusion probabilistic model that estimates complete radio maps from either sparse RSS fragments or transmitter locations. The paper formulates the problem as a conditional generation task, designs separate encoders for the two condition types, and proposes an environment-aware fragment-selection heuristic based on obstacle density. Training and test data are generated with the Wireless InSite ray tracer for indoor 60 GHz and outdoor 3.7 GHz scenarios. The experiments report that RM-Gen reaches 96.85% indoor and 89.54% outdoor accuracy at ETR=0.10 with 5% RSS fragments, outperforming cGAN and pix2pix baselines (Tables II and III). The paper also shows visual map comparisons and two downstream use cases.","tokens_in":18748,"tokens_out":5927,"duration_ms":53591,"significance":"If the reported evaluations are reproducible, the paper provides the first demonstration that conditional diffusion models can estimate radio maps from sparse measurements and Tx coordinates, with a practical fragment-selection algorithm. Strengths include the standard DDPM training objective (Eq. 14), explicit pseudocode for training, sampling, and fragment selection, and the construction of two sizeable synthetic datasets (~30k and ~26k maps). The environment-aware selection method is a reasonable heuristic and is shown to outperform random selection. However, the central claim depends on a conditioning mechanism whose current description is underspecified, and the evaluation protocol lacks a defined train/test split and exact metric formula. These issues must be resolved before the empirical claims can be accepted.","major_comments":[{"comment":"The RSS-fragment encoder described by Eqs. (15)-(18) concatenates flattened RSS values only; it does not include fragment coordinates, a spatial mask, or any positional embedding in e_theta(c). Such a condition vector is permutation-invariant across fragments and cannot align a measured fragment to its location in the N by N map. Table II reports 96.85% accuracy at 5% fragments, which cannot be explained by the described architecture unless the experiment supplies location information implicitly (e.g., fixed fragment positions or a hidden coordinate channel). Please state precisely what input the encoder receives and, if the implementation used positional information, include that in the description.","section":"IV.B, Eqs. (15)-(18)"},{"comment":"The paper does not specify how the approximately 30,000 (RM-In) and 26,000 (RM-Out) radio maps are split into training, validation, and test sets. Without this split, it is unclear whether the reported accuracies reflect memorization of seen scenario configurations or generalization to unseen AP locations and environments. Please describe the split criterion (e.g., by random map, by AP position, or by environment layout) and confirm that test maps are not used in training.","section":"V, experimental setup"},{"comment":"The Error Tolerance Rate is defined only verbally as \"the maximum allowable percentage difference\" and \"a performance difference ratio of 10% is acceptable.\" The exact formula is missing: it is not stated whether accuracy is the fraction of grid points with relative error below the threshold, the mean over the map, or a normalized RSS error. Since all headline numbers (Tables II-III, Figs. 6-7, 12) depend on this metric, please give the precise definition, including how negative RSS values (in dBm) are handled.","section":"V.B, ETR definition"},{"comment":"The cGAN and pix2pix baselines are described at a high level, but no hyperparameters (learning rates, number of epochs, optimizer settings, or discriminator architecture beyond \"MLP-based\") are given. Without evidence that these baselines were tuned to a comparable degree, the reported 20%+ margin over cGAN and about 10% over pix2pix in Table II may reflect undertuned baselines. Please provide baseline training details and report mean and standard deviation over multiple random seeds.","section":"V.D, baseline comparison"},{"comment":"The comparison and all conclusions are grounded in ray-tracing-generated ground-truth maps. The two citations [47], [48] support general closeness of ray tracing to measurements, but the specific indoor and outdoor layouts in this paper are not validated against real RSS measurements. Since both RM-Gen and the baselines are trained and tested on the same simulator, the relative ranking may survive, but the absolute accuracy figures may not transfer to physical deployments; please add a discussion of this limitation and, if possible, a small real-data validation.","section":"IV.D, data collection"}],"minor_comments":[{"comment":"The word \"popouted\" in the sentence following Algorithm 3 appears to be a typo; it should be \"output\" or \"selected.\"","section":"IV.C, Algorithm 3"},{"comment":"The phrase \"out door\" in the comparison paragraph should be \"outdoor.\"","section":"V.D"},{"comment":"In Eq. (19), the text states \"where L denotes the concatenation,\" which appears to be a typo for the concatenation symbol (e.g., the big oplus used in Eq. 16) or should be written explicitly as \"where the big operation denotes concatenation.\"","section":"IV.B, Eq. (19)"},{"comment":"The reference to \"Fig. 11(b)\" for the BS/UAV use case is likely incorrect; the use-case figure appears to be Fig. 13(b), since Fig. 11 shows an outdoor radio map comparison.","section":"V.E"},{"comment":"The abstract claims \"over 95% accuracy\" for both 60 GHz and sub-6GHz networks, but the reported outdoor accuracy is around 90% at ETR=0.10 (Fig. 7(b), Table II); the conclusion's wording (95% indoor, 90% outdoor) is more accurate and should be used consistently.","section":"Abstract and V.B"},{"comment":"Table I uses \"T x\" while equations use \"Tx\"; please unify the notation. Also, Figs. 6, 7, and 12 use \"accuracy\" without a formal definition; once the ETR formula is provided, define \"accuracy\" in the figure captions.","section":"Notation and figures"}],"recommendation":"major_revision","confidential_remarks":"This submission is an extension of the authors' IFIP Networking 2024 paper [45], adding the environment-aware selection method and a second condition type. The topic fits the journal's scope, and the diffusion-model application to radio maps is timely. However, the evaluation reporting is not yet at the standard required for the central claims: the RSS-fragment encoder as written lacks positional information, and the train/test split and ETR metric are unspecified. I recommend requesting a revised version that resolves the conditioning mechanism and reports the missing evaluation details, rather than rejecting the work, since the core idea is promising and the identified issues appear addressable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a straightforward application of conditional DDPM to radio map estimation, with a sensible environment-aware heuristic for picking sparse RSS fragments. Second, the central mechanism as written cannot explain the results: the fragment encoder flattens and concatenates RSS values only, with no coordinates or spatial mask, so the model has no way to know where a fragment belongs. At 5% fragments, the reported 96.85% accuracy is not credible under the described architecture.\n\nWhat is genuinely in the paper: a clean DDPM formulation, two condition types (RSS fragments and Tx locations), a simple obstacle-density-based fragment selection rule that beats random selection, and a head-to-head comparison with cGAN and pix2pix on two large ray-traced datasets. The Tx-location-only setting is a practical scenario worth exploring. The authors also cite their own prior RM-Gen work, so the incremental nature is acknowledged even if the 'first endeavor' claim is overstated.\n\nThe soft spots are real. The missing fragment-coordinate information is load-bearing; either the implementation silently feeds positions and the paper omits it, or the evaluation protocol fixes fragment locations in a way that leaks the answer. Either way, the text does not support the claim. On top of that, there is no train/test split described, no error bars or multiple runs, no code or data release, and the ETR metric is never precisely defined. The baselines may be undertuned—no hyperparameters are given. And the abstract's 'over 95% for both' is contradicted by the ~90% outdoor numbers in the body. The reliance on Wireless InSite for both training and evaluation is a threat, though less central than the conditioning gap.\n\nI would not publish this as is. It deserves a serious referee, because the problem is real and the approach would be valuable if the missing coordinate channel exists and the evaluation is clean. But the authors need to fix the description of the conditioning mechanism, report the split, and share code or detailed experimental settings. Send it back for major revision.","headline":"A plausible conditional-DDPM radio map generator whose reported accuracy is unsupported because the described fragment encoder never sees where fragments are.","tokens_in":19294,"tokens_out":4242,"would_cite":false,"duration_ms":40784,"reading_group":"no","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A conditional diffusion model turns sparse signal samples and transmitter locations into accurate radio maps.","keywords":["radio map estimation","denoising diffusion probabilistic models","conditional diffusion model","generative wireless networks","millimeter-wave WLAN","sub-6 GHz networks","RSS fragment selection","ray tracing"],"falsifier":"Run the same RM-Gen pipeline with real field measurements in a 60 GHz indoor office and a 3.7 GHz outdoor site, using exactly 5%, 10%, and 15% RSS fragments and the same ETR=0.10 rule; if accuracy falls below the reported 95% indoor and 90% outdoor levels, or if pix2pix matches or beats RM-Gen, the central claim fails. Before that, a cheaper check is to inspect the dataset split: if held-out test maps share AP positions, layouts, or fragment locations with training maps, the reported numbers are inflated.","tokens_in":18271,"feed_emoji":"📶","tokens_out":8039,"duration_ms":58228,"temperature":0.7,"pith_summary":"This paper sets out to show that a conditional denoising diffusion probabilistic model, called RM-Gen, can reconstruct a complete radio map from very little measured information: a handful of received-signal-strength fragments or, in the simplest case, only the coordinates of the transmitters. The authors report accuracy above 95% for indoor 60 GHz WLANs and about 90% for outdoor sub-6 GHz cellular environments when RSS fragments make up 5–15% of the map, with RM-Gen outperforming cGAN and pix2pix baselines by roughly 10–20 percentage points. The practical stake is that radio maps, which are normally obtained through expensive measurement campaigns or slow ray-tracing simulations, could be generated cheaply enough for routine network planning, digital-twin modeling, and drone coverage optimization.","feed_headline":"Rebuilds radio maps from 5% signal samples via diffusion model","feed_subtitle":"Indoor 60 GHz maps reach over 95 percent accuracy with sparse RSS fragments, beating GAN and pix2pix baselines.","key_machinery":"The load-bearing mechanism is the conditional denoising diffusion process. Given a ground-truth radio map $x_0$, a schedule of Gaussian noises produces a noisy map $x_t = \\sqrt{\\bar{\\alpha}_t} x_0 + \\sqrt{1-\\bar{\\alpha}_t} \\epsilon$; the model learns a denoising function $\\epsilon_\\theta(x_t, t | e_\\theta(c))$ implemented by a U-Net, a convolutional encoder-decoder network, conditioned on $e_\\theta(c)$, an encoding of the available information $c$. For RSS fragments, $e_\\theta$ flattens and concatenates fragment vectors and passes them through a three-layer MLP; for transmitter locations, it concatenates learned embeddings of each coordinate. Because the loss is simply the mean squared error between predicted and actual noise, the same training loop works for either condition. The environment-aware fragment selection algorithm divides the scenario into subareas, counts obstacle density in each, and chooses fragments centered in the densest subareas, which are presumed to carry the most propagation information.","core_discovery":"The central claim is that radio map estimation can be posed as a conditional generation problem and solved with a denoising diffusion probabilistic model: a forward process adds Gaussian noise to a ground-truth map over 400 time steps, and a U-Net trained to remove that noise, guided by an encoded condition, generates the map in reverse. Two kinds of conditions work: flattened RSS fragments processed by an MLP encoder, and embedded transmitter coordinates. The paper further claims that the choice of which fragments to measure matters, and that sorting candidate locations by local obstacle density yields accuracy close to human expert selection while random selection lags far behind. On the authors' ray-traced datasets, RM-Gen achieves 96.85–98.09% accuracy on indoor mmWave maps and 89.54–90.97% on outdoor sub-6 GHz maps at ETR=0.10 with RSS fragments, and 81.95% indoor and 74.26% outdoor with Tx locations alone.","pith_inferences":["The paper leaves implicit that its ground truth is ray-traced rather than measured; a real-world deployment could show a smaller accuracy gap over baselines if ray tracing misses material, weather, or motion effects.","The paper never states the train/test split, so a necessary check is whether held-out scenarios differ in AP locations, layouts, and fragment placements; if they overlap, the reported accuracies could be optimistic.","A natural extension the paper only mentions in passing is fusing both conditions, RSS fragments plus transmitter coordinates, through a shared attention or MLP module; this could lift the weaker Tx-location-only results.","The ETR=0.10 metric treats any pixel within 10% of ground truth as correct; tasks like beam management may require stricter spatial accuracy, so the headline percentages are not end-to-end task performance."],"forward_implications":["Radio map generation no longer requires a full environmental model: sparse RSS fragments alone are enough for high-accuracy reconstruction in the tested indoor scenario.","Planning can happen before deployment: with only planned transmitter positions as input, RM-Gen still produces usable maps, supporting access point placement and UAV trajectory design without measurements.","The environment-aware selection rule gives network operators a no-expert alternative to manual fragment placement, closing most of the gap to human-labeled selection.","At the tested fragment fractions, RM-Gen outperforms cGAN and pix2pix in every reported configuration, suggesting that iterative denoising is better suited than adversarial training to reconstructing fine-grained signal structure."],"supporting_citations":[{"why":"It supplies the denoising diffusion probabilistic model and its training objective that RM-Gen adapts to radio maps.","marker":"[28]"},{"why":"It provides the U-Net architecture used as the noise-prediction backbone in the reverse diffusion process.","marker":"[46]"},{"why":"It defines the conditional GAN baseline that RM-Gen is compared against for both condition types.","marker":"[49]"},{"why":"It defines the pix2pix image-to-image translation baseline with its combined adversarial and L1 loss.","marker":"[50]"},{"why":"It supports the paper's assumption that ray-traced signal profiles approximate real measurements in wireless scenarios.","marker":"[47]"},{"why":"It gives the direct comparison of ray-tracing simulations with millimeter-wave channel sounding used to justify the ground-truth dataset.","marker":"[48]"}],"fun_headline_variants":["Diffusion model rebuilds radio maps from 5% signal samples","Radio map estimation via diffusion from sparse sample sets","Sparse RSS fragments to radio maps with conditional diffusion","Environment-aware diffusion yields 95%+ accurate radio maps","Diffusion beats GAN and pix2pix for radio map generation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracy numbers rest on two linked assumptions: that ray-traced maps stand in for real-world received signal strength measurements, and that training and test maps come from the same distribution without leakage.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model rebuilds radio maps from 5% signal samples","Radio map estimation via diffusion from sparse sample sets","Sparse RSS fragments to radio maps with conditional diffusion","Environment-aware diffusion yields 95%+ accurate radio maps","Diffusion beats GAN and pix2pix for radio map generation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001048,"raw_usage":{"total_tokens":4418,"prompt_tokens":971,"completion_tokens":3447,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":3364}},"tokens_in":587,"tokens_out":3447,"duration_ms":80168,"temperature":1.0,"reasoning_tokens":3364,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:56:35.665688+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same RM-Gen pipeline with real field measurements in a 60 GHz indoor office and a 3.7 GHz outdoor site, using exactly 5%, 10%, and 15% RSS fragments and the same ETR=0.10 rule; if accuracy falls below the reported 95% indoor and 90% outdoor levels, or if pix2pix matches or beats RM-Gen, the central claim fails. Before that, a cheaper check is to inspect the dataset split: if held-out test maps share AP positions, layouts, or fragment locations with training maps, the reported numbers are inflated.","supporting_citations":[{"cited_title":"6G wireless channel measurements and models: Trends and challenges,","cited_arxiv_id":null,"evidence_quote":"It supports the paper's assumption that ray-traced signal profiles approximate real measurements in wireless scenarios."},{"cited_title":"Comparison of ray tracing simulations and millimeter wave channel sounding mea- surements,","cited_arxiv_id":null,"evidence_quote":"It gives the direct comparison of ray-tracing simulations with millimeter-wave channel sounding used to justify the ground-truth dataset."}],"review_version":1}