{"id":"6b051232-a3ce-45ae-ae2e-3527cef38807","arxiv_id":"2508.09140","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"On the RadioMapSeer benchmark, RadioMamba reports NMSE 0.0050 versus 0.0072 for RadioDiff, with 28 ms inference and 8.6M parameters.","lead":"RadioMamba is a neural network that builds radio coverage maps from building layouts and transmitter positions by combining a global-context Mamba branch with a local convolutional branch inside a U-Net. It reports higher accuracy than prior models while running about 20 times faster and using only 2.9 percent of the parameters of the diffusion baseline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Composite-loss confound: RadioMamba's own Table V shows the tuned SSIM+gradient loss explains a 33% NMSE improvement, larger than the 30.5% gap over RadioDiff, yet baselines were not trained with the same loss.","rationale":"The reader's weakest_assumption correctly flags baseline comparability, but it is underspecified. The paper's own Table V sharpens the issue: full-loss RadioMamba (NMSE 0.0050) beats L1+MSE-only RadioMamba (0.0075) by 33%, which is larger than the 30.5% margin over RadioDiff (0.0072). This means the reported accuracy gain is plausibly a loss-tuning effect rather than a Mamba-UNet architecture effect, unless the baselines used the identical loss. Table IV's depthwise-vs-standard convolution ablation and Table V's loss ablation are useful internal controls, and the efficiency numbers in Table III are less affected by this confound, so the paper still has value. However, the central claim that the hybrid architecture breaks the accuracy-efficiency trade-off is not currently supported by a controlled comparison. This is an evidential gap, not a demonstrated falsehood, so the reader's CONDITIONAL verdict should stand unchanged. I would not escalate to REJECT because the reported numbers may be reproducible; the paper simply needs a matched-loss, matched-hardware baseline evaluation.","tokens_in":18147,"tokens_out":7011,"duration_ms":79764,"concrete_test":"Retrain RadioDiff and RadioUNet on RadioMapSeer using RadioMamba's exact composite loss (Eq. 26) with identical preprocessing, data split, and training schedule, then evaluate SRM/DRM NMSE and inference time on the same A40 GPU with the same batch size. If RadioDiff's NMSE falls to or below RadioMamba's 0.0050 within error bars, the architecture-specific SOTA claim fails; if RadioDiff stays near 0.0072, the loss confound is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim depends on an uncontrolled loss-function comparison. Section V-A4 defines a composite loss (L1+MSE+SSIM+gradient, Eq. 26) with weights tuned to the benchmark. Table V reports that RadioMamba trained with only L1+MSE reaches SRM NMSE 0.0075, slightly worse than RadioDiff's reported 0.0072; the full loss improves it to 0.0050, a 33% relative reduction. The claimed margin over RadioDiff is 30.5%, so the loss function alone can account for more than the entire accuracy advantage. Section V-A2 lists baselines but never states they were retrained with the same loss, preprocessing, or schedule; Section V-A3 gives implementation details only for RadioMamba. Thus the headline result conflates architectural benefit with loss engineering. If RadioDiff also benefits from SSIM/gradient terms, the accuracy claim may vanish even though the efficiency and parameter-count advantages (Table III) would remain.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RadioMamba, a hybrid Mamba-UNet architecture for sampling-free radio map construction. The core MambaConvBlock combines a bidirectional Mamba branch meant to capture long-range spatial dependencies with a depthwise-separable convolutional branch that extracts local features. On the RadioMapSeer benchmark, the authors report NMSE 0.0050 for static radio map construction and 0.0063 for dynamic radio map construction, outperforming RadioUNet, RME-GAN, and the diffusion-based RadioDiff, while inference time is 0.028 s and parameter count is 8.6M. The paper includes ablations of the block components, the use of depthwise separable convolutions, and the composite loss function.","tokens_in":18317,"tokens_out":3394,"duration_ms":37127,"significance":"If the accuracy comparison is fair, the paper would make a strong practical contribution by showing that a single-pass, lightweight model can match or beat an iterative diffusion model at a fraction of the computational cost. The architectural motivation is sound, and the ablations support the usefulness of both branches and of depthwise separable convolutions. The efficiency and parameter-count advantages are likely robust. However, the headline accuracy claim is currently threatened by the loss-function confound and by the lack of a controlled baseline protocol, so the significance of the accuracy result cannot be fully assessed as presented.","major_comments":[{"comment":"The composite loss is a major confound. Training with only L1+MSE yields NMSE 0.0075 on SRM, which is worse than RadioDiff's reported 0.0072; the full loss improves NMSE to 0.0050, a 33% relative improvement. The paper's headline margin over RadioDiff is 30.5%, so the loss function alone can account for the entire accuracy gap. Since Section V-A2 does not state that baselines were trained with the same loss, the central 'higher accuracy' claim is not established. Please retrain RadioDiff (and other baselines) with the same composite loss, or otherwise demonstrate that the loss does not preferentially benefit RadioMamba.","section":"V-A4, Eq. (26), Table V"},{"comment":"The baseline comparison is not controlled. The paper never states whether RadioUNet, RME-GAN, and RadioDiff were retrained under the same preprocessing, loss, optimizer schedule, hardware, and batching, or whether the numbers are copied from prior publications. In particular, RadioDiff's 0.5535 s inference time may come from a different GPU or batch configuration; without a common timing protocol the 'nearly 20 times faster' claim is not convincing. Please specify the baseline evaluation protocol and, ideally, rerun all baselines in the same environment.","section":"V-A2, V-A3"},{"comment":"No error bars, confidence intervals, or repeated seeds are reported. Accuracy metrics like NMSE 0.0050 vs 0.0072 are point estimates from, presumably, a single run. Given the small absolute differences, these results do not establish statistical significance. Please report mean and standard deviation over at least three seeds and make the code available to support reproducibility.","section":"Tables II and V"}],"minor_comments":[{"comment":"The abbreviations 'NMSA' and 'RMSA' in the radar chart are not defined; please spell them out in the caption or use the standard NMSE and RMSE names.","section":"Fig. 1"},{"comment":"Please clarify whether the dynamic obstacle channel is zero-filled for the SRM task or omitted from the input tensor, since the current text says it 'may be zero-filled or omitted.'","section":"III-A"},{"comment":"The acknowledgment of the anisotropy of raster-scan flattening is a useful caveat; please also mention it in the conclusions as a known limitation.","section":"IV-B2"},{"comment":"State explicitly in the table footnote that the percentages indicate improvement relative to RadioDiff, to avoid ambiguity about the reference point.","section":"Table III"},{"comment":"The data split of 550/50/100 maps sums to 700 total maps; please confirm that this follows the original RadioMapSeer split and cite the corresponding benchmark documentation.","section":"V-A1"}],"recommendation":"major_revision","confidential_remarks":"The RadioDiff baseline appears to come from the same research group. It would strengthen the paper to state in the text whether the RadioDiff authors provided the original configuration or whether the numbers were reproduced independently. Given the reproducibility-related revisions requested, code release would be particularly valuable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: RadioMamba is a competent engineering contribution. The hybrid Mamba-UNet is not architecturally new, but the adaptation to radio map construction is solid, and the ablations genuinely support the design choices. The efficiency results — 28 ms inference, 8.6M parameters, 808 MB memory versus RadioDiff's 553 ms, 297.74M, 2067 MB — are striking and likely to survive scrutiny, because they do not depend on the loss function. That part is the real news: a single-pass model with global receptive field can approach or beat diffusion accuracy at a fraction of the cost.\n\nThe soft spot is the accuracy claim. The stress-test note is correct and important: Table V shows that switching from L1+MSE alone to the full composite loss improves SRM NMSE from 0.0075 to 0.0050, a 33% improvement. The claimed 30.5% gain over RadioDiff's 0.0072 is smaller than the loss effect. The paper never states whether baselines were retrained with the same loss, preprocessing, or schedule; the details in Section V-A3 are only for RadioMamba. So the accuracy margin over RadioDiff could be mostly loss engineering, not architecture. That does not kill the paper, but it means the headline \"higher accuracy than diffusion models\" is not yet established. The authors should retrain RadioUNet, RME-GAN, and RadioDiff under identical training conditions and loss, or at minimum report the baselines with the same loss. Also missing: error bars or repeated seeds, code release, and any baseline training/timing details. These are all fixable before publication.\n\nThe citation pattern is fine — RadioDiff is from the same group, but that is the benchmark SOTA and the comparison is central. The self-citation is not a flaw in itself.\n\nWho is this for? Wireless researchers working on radio map estimation or 6G digital twins, and anyone interested in efficient alternatives to diffusion models for image-to-image tasks. It deserves a serious referee: the question of whether a hybrid Mamba-CNN can replace diffusion in this domain is important, and the authors have done enough work to justify a careful look. I would send it to peer review with a strong request that the accuracy comparison be made fair and that code be released.","headline":"A useful engineering paper whose headline accuracy claim is undercut by an uncontrolled loss-function comparison, but whose efficiency gains are real and worth reviewing.","tokens_in":18874,"tokens_out":1251,"would_cite":false,"duration_ms":17124,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"RadioMamba claims a single-pass U-Net with a Mamba branch can build radio maps more accurately than iterative diffusion models while running nearly 20 times faster and using 2.9 percent of the parameters.","keywords":["radio map construction","Mamba","state space models","U-Net","6G networks","radio propagation modeling","lightweight deep learning","diffusion models"],"falsifier":"Run RadioUNet, RME-GAN, and RadioDiff on the RadioMapSeer test split under exactly the conditions used for RadioMamba—same normalization, same composite loss, same optimizer schedule, same batch size, and same GPU—and compare per-map inference time and NMSE. The central claim fails if RadioDiff then matches or beats RadioMamba's NMSE, or if its inference time drops to RadioMamba's level.","tokens_in":17941,"feed_emoji":"📡","tokens_out":6389,"duration_ms":67567,"temperature":0.7,"pith_summary":"Radio map construction—predicting signal strength across a city from building layouts and transmitter positions—is caught between accurate but slow generative models and fast but less accurate convolutional ones. The paper argues this trade-off is architectural: convolutional receptive fields are too local to capture distant obstacles and reflections, while global-context models such as Transformers are too costly, and iterative diffusion models are accurate but slow. RadioMamba is a single-pass U-Net whose building block runs a Mamba state-space branch in parallel with a lightweight convolutional branch, giving every pixel access to global context at linear cost while preserving local detail. On the RadioMapSeer benchmark, the paper reports the best NMSE, RMSE, SSIM, and PSNR among compared methods, including RadioDiff, while inference takes 28 ms versus 553 ms and the model has 8.6 million parameters versus 297.74 million. If these numbers hold, the accuracy-efficiency trade-off in learned radio maps would no longer be a binding constraint for real-time use.","feed_headline":"Mamba-UNet radio maps beat diffusion at 1/20 the latency","feed_subtitle":"RadioMamba reports higher accuracy with 8.6M parameters versus 297.7M, pointing to real-time 6G deployment.","key_machinery":"The load-bearing component is the MambaConvBlock, a parallel two-branch module used at every encoder and decoder stage of a U-Net. One branch, SS2D-Mamba, flattens the feature map into a sequence, runs a selective state-space model in both forward and backward directions (a bidirectional raster scan), and reshapes the result, so each pixel can aggregate information from the whole map in linear time; the other branch is a residual depthwise-separable convolution that captures local edges and fine detail. The outputs are combined by element-wise addition, and the paper argues that the convolutional branch's isotropic inductive bias also compensates for the anisotropic scan of the Mamba branch. These blocks are embedded in a standard U-Net with skip connections, and the model is trained with a composite loss of L1, MSE, SSIM, and gradient terms.","core_discovery":"The central claim is that a hybrid Mamba-convolutional U-Net can outperform the current state of the art, including diffusion-based RadioDiff, on both accuracy and efficiency for sampling-free radio map construction. The paper attributes this to physics: pathloss at any point depends on long-range spatial structure such as distant buildings casting shadows, so a model needs a global receptive field, but it also needs sharp local boundaries around obstacles. RadioMamba's MambaConvBlock provides both: a bidirectional Mamba scan over the rasterized feature map models global dependencies in linear time, while a depthwise-separable convolutional branch handles local texture, and the two are fused by element-wise addition. On the static-map task the reported NMSE is 0.0050 versus 0.0072 for RadioDiff, and on the dynamic-map task 0.0063 versus 0.0090, with inference time of 0.0280 s versus 0.5535 s and 8.6 million versus 297.74 million parameters.","pith_inferences":["Beyond the paper: the reported speedup and parameter ratio would be decisive only if baselines were retrained or re-measured under identical conditions; a head-to-head rerun is the test that would settle the comparison.","Beyond the paper: the architecture's anisotropic raster scan may be a limitation, but the convolutional branch's isotropic bias may mask it; trying a space-filling-curve scan could test whether alignment with isotropic propagation further improves accuracy.","Beyond the paper: the same MambaConvBlock could plausibly transfer to other physics-driven image-to-image tasks, such as indoor coverage maps or channel knowledge maps, wherever long-range spatial dependencies dominate."],"forward_implications":["Radio maps can be produced at 28 ms per 256x256 map, within the latency budget for real-time network control loops and UAV trajectory planning.","The 97 percent parameter reduction relative to RadioDiff makes edge-device deployment a realistic option for the same accuracy class.","Because accuracy improves while inference is single-pass, iterative diffusion sampling is not required to reach state-of-the-art results on this benchmark.","The dynamic-obstacle channel shows the same gains, so the method extends beyond static city maps to changing environments with transient blockers."],"supporting_citations":[{"why":"Supplies the RadioMapSeer dataset and the CNN baseline whose efficiency RadioMamba extends.","marker":"[29]"},{"why":"Provides the GAN-based baseline for accuracy and inference-time comparison.","marker":"[30]"},{"why":"Is the diffusion-model state of the art that RadioMamba claims to beat in accuracy while running faster with fewer parameters.","marker":"[31]"},{"why":"Introduces structured state space models that make efficient global sequence modeling possible.","marker":"[35]"},{"why":"Defines the selective state-space model that gives the Mamba branch its linear-complexity global receptive field.","marker":"[36]"},{"why":"Provides the U-Net encoder-decoder skeleton with skip connections that RadioMamba fills with MambaConvBlocks.","marker":"[50]"},{"why":"Shows how to adapt Mamba to 2D image data with bidirectional scans, the basis of SS2D-Mamba.","marker":"[65]"}],"fun_headline_variants":["RadioMamba: 20x faster, 2.9% params, better radio maps","Mamba-UNet radio maps beat diffusion at 20x speed","Hybrid Mamba-UNet: linear-time global context for radio maps","RadioMamba: accuracy up, compute down 20x for radio maps","Break trade-off: RadioMamba does radio maps faster and better"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline numbers assume that the baselines, especially RadioDiff, were evaluated under the same preprocessing, loss, training schedule, and hardware as RadioMamba; if RadioDiff's 553 ms timing came from a different setup, the claimed 20x speedup and accuracy advantage would not be a like-for-like comparison.","fun_headline_variants_meta":{"raw":{"variants":["RadioMamba: 20x faster, 2.9% params, better radio maps","Mamba-UNet radio maps beat diffusion at 20x speed","Hybrid Mamba-UNet: linear-time global context for radio maps","RadioMamba: accuracy up, compute down 20x for radio maps","Break trade-off: RadioMamba does radio maps faster and better"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000269,"raw_usage":{"total_tokens":1624,"prompt_tokens":949,"completion_tokens":675,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":573}},"tokens_in":565,"tokens_out":675,"duration_ms":6922,"temperature":1.0,"reasoning_tokens":573,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:33:25.099779+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run RadioUNet, RME-GAN, and RadioDiff on the RadioMapSeer test split under exactly the conditions used for RadioMamba—same normalization, same composite loss, same optimizer schedule, same batch size, and same GPU—and compare per-map inference time and NMSE. The central claim fails if RadioDiff then matches or beats RadioMamba's NMSE, or if its inference time drops to RadioMamba's level.","supporting_citations":[{"cited_title":"RME-GAN: A learning framework for radio map estimation based on conditional generative adversarial network,","cited_arxiv_id":null,"evidence_quote":"Provides the GAN-based baseline for accuracy and inference-time comparison."},{"cited_title":"RadioDiff: An effective generative diffusion model for sampling- free dynamic radio map construction,","cited_arxiv_id":null,"evidence_quote":"Is the diffusion-model state of the art that RadioMamba claims to beat in accuracy while running faster with fewer parameters."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the U-Net encoder-decoder skeleton with skip connections that RadioMamba fills with MambaConvBlocks."},{"cited_title":"Vmamba: Visual state space model,","cited_arxiv_id":null,"evidence_quote":"Shows how to adapt Mamba to 2D image data with bidirectional scans, the basis of SS2D-Mamba."}],"review_version":1}