{"id":"6f212438-0477-4941-b712-ed8bb05e357d","arxiv_id":"2607.17655","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DMSNet, a coarse-to-fine dual-band neural network, jointly estimates target count and range/velocity/angle from 3.5/28 GHz ISAC echoes, raising count accuracy to 91.74% and cutting median parameter errors by 56–82% in simulation.","lead":"A new neural network called DMSNet combines echoes from 3.5 GHz and 28 GHz radio bands to count objects and estimate their range, speed, and angle, reporting large accuracy gains over classic baselines in simulation. The payoff, if it survives real-world testing, is cheaper and more reliable multi-target sensing for 6G networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-band complementarity is asserted, not demonstrated: the Sionna RT dataset's per-band RCS statistics are unspecified, so the count-accuracy gain may be an artifact of synthetic ρ_b,l rather than a real multi-band ISAC effect.","rationale":"The reader's weakest_assumption correctly identified the same load-bearing concern: the realism of the simulated cross-band complementarity. Our stress test agrees. We considered other concerns (no variance, runtime claim, baseline reimplementation) but they are addressable and less fundamental. The no-variance issue is important but would not change the sign of the effect if true; the runtime claim is a wording issue; baseline reimplementation could be checked by code release. The RCS realism issue is load-bearing because the paper's novelty is cross-band fusion, and without realistic complementarity the result is a synthetic demonstration. We therefore recommend no change to the conditional verdict: acceptance should require release of the dataset generation code/parameters and a demonstration that the simulated band-dependent visibility statistics are physically plausible (or a shared-RCS ablation).","tokens_in":9463,"tokens_out":6037,"duration_ms":56307,"concrete_test":"Rerun the enumeration module (and ideally the full pipeline) on a variant of the test set where each target's RCS is drawn once per scene and shared across both bands (ρ_{h,l}=ρ_{l,l} per target; i.e., no band-dependent visibility), while keeping all other simulation settings identical. If the count-accuracy gap between DMSNet and the high-only baseline (currently 91.81% vs 84.23%, Table III) collapses or reverses, the claimed cross-band fusion gain is an artifact of the unspecified per-band RCS sampling. Report the per-target band-SNR correlation as well.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — DMSNet's learned cross-band fusion beats handcrafted fusion — depends on Eq. (1)'s frequency-selective complex scattering coefficient ρ_{b,l}, which is invoked to justify that 'a target weakly visible in one band may still be detectable in another band.' The simulation section (§IV-A) states only that the dataset was generated in a Sionna RT digital twin of the BUPT campus; it does not specify how ρ_{b,l} (or the resulting per-band target RCS) is obtained. Sionna RT is a radio-propagation ray tracer; without an explicit target RCS model, it does not by itself produce realistic frequency-dependent target scattering. If ρ_{b,l} are drawn independently per band, cross-band complementarity is manufactured and DMSNet's counting gain is trivial. This matters because Table III shows the fusion benefit is concentrated in counting (high-only 84.23% vs DMSNet 91.81%), whereas parameter accuracy is dominated by the fine refinement module (coarse-only P50 range 2.1192 m vs DMSNet 0.0110 m) and high-only P50 range (0.0109 m) is already equal to DMSNet (0.0110 m). Thus the entire qualitative claim of cross-band advantage rests on unvalidated RCS statistics, and the paper provides no comparison with measured dual-band RCS data (e.g., from [9]) to support the simulated complementarity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DMSNet, a neural-network architecture for dual-band (3.5 GHz / 28 GHz) ISAC multi-target sensing. DMSNet jointly estimates the number of targets and their range, velocity, and angle by cascading three modules: a target-number module based on 3D CNN encoders and multi-head cross-band attention; a coarse parameter module that casts parameter estimation as bin classification via a transformer decoder; and a fine module that extracts local 256×256×256 FFT patches around coarse centers and regresses residual offsets. Training is module-wise, with a count loss, a Hungarian-matched coarse loss, and a combined heatmap/Smooth-L1 fine loss. The method is evaluated on a synthetic Sionna RT digital-twin dataset of the BUPT campus with up to five UAV targets, against CA-CFAR-based, CFAR-based, and learning-based baselines. The paper reports improved count accuracy (91.74% vs. 89.01%), reduced median absolute range/velocity/angle errors by 82.2%, 56.9%, and 73.2%, and a runtime of 23.7 ms, which is 68.7% faster than the fastest conventional dual-band pipeline (CP-based).","tokens_in":9738,"tokens_out":5685,"duration_ms":57551,"significance":"If the results hold, DMSNet is a useful step toward learned cross-band fusion in multi-band ISAC. The architecture is clearly specified, the three-module cascade is well motivated, and the ablation study isolates the contributions of cross-band fusion and fine refinement. The evaluation is scoped to simulation and does not overclaim real-world performance. However, the central claim that learned cross-band fusion exploits physical complementarity rests on the realism of the simulated frequency-selective target scattering, and the reported gains are single-run numbers without uncertainty quantification. The runtime claim in the abstract is stronger than the comparison in Table II supports. These issues are fixable but require additional evidence.","major_comments":[{"comment":"The load-bearing premise is the frequency-selective scattering coefficient ρ_{b,l} in Eq. (1), justified by the statement that a target weakly visible in one band may still be detectable in another band. Section IV-A describes the Sionna RT campus digital twin but does not specify how ρ_{b,l} or the resulting per-band RCS ξ_{b,l} is assigned to UAV targets. Sionna RT is a radio-propagation ray tracer; it does not by itself supply a UAV RCS model. If ρ_{b,l} is drawn independently per band, cross-band complementarity is manufactured, and Table III shows that the fusion benefit is concentrated in counting (High-only 84.23% vs. DMSNet 91.81%), while parameter medians are essentially equal to High-only (e.g., range P50 0.0109 m vs. 0.0110 m). Please report the RCS model, per-band target SNR statistics, the correlation between bands, and a sensitivity study, or use measured dual-band RCS data","section":"§IV-A and Eq. (1)"},{"comment":"All evaluation numbers are single-run, with no seeds, error bars, or confidence intervals. Headline statements such as the 82.2%, 56.9%, and 73.2% median-error reductions and the count-accuracy difference (91.74% vs. 89.01%) are therefore not statistically grounded. The discrepancy between DMSNet count accuracy in Table I (0.9174) and Table III (0.9181) at the same SNR suggests run-to-run variation. Please provide results over multiple training seeds with mean and standard deviation or confidence intervals, and a significance test for the count-accuracy differences.","section":"§IV-A, Tables I–III, Fig. 4"},{"comment":"The abstract claims DMSNet 'reduces runtime by 68.7% relative to the fastest existing dual-band ISAC sensing method,' but Table II shows CSIYOLO-based runs in 9.740 ms, faster than DMSNet's 23.697 ms. The 68.7% reduction is only relative to the fastest conventional pipeline, CP-based (75.665 ms). Please correct the abstract and any similarly worded claims, or justify why CSIYOLO-based is excluded from the runtime comparison.","section":"§IV-A and §IV-D"},{"comment":"The baselines are reimplementations, but implementation details are insufficient for reproducibility. In particular, 'CSIYOLO-based' is described only as a dual-band adaptation with aligned and fused band-specific features; no architecture details, training schedule, or hyperparameters are given. The signal-processing baselines [3]–[6] are also reimplemented without stating how carefully their parameters were tuned. If the baselines are under-tuned, the reported error reductions are inflated. Please provide complete baseline configurations and, ideally, release code or pseudo-code for the evaluation pipeline.","section":"§IV-A, §IV-C, §IV-D"},{"comment":"The parameter-error CDFs do not specify how predictions are matched to ground truth when the target count is estimated incorrectly. If errors are computed only on Hungarian-matched predicted targets, misses and false positives are not penalized, which can artificially improve parameter metrics. Please define the exact matching and scoring protocol, including what happens when the estimated count differs from the true count.","section":"§IV-C"}],"minor_comments":[{"comment":"The abstract and the conclusion repeat the same numbers; the conclusion would benefit from one sentence on limitations, especially the dependence on simulated scattering models.","section":"General"},{"comment":"'CandR' is a LaTeX rendering issue; also, the notation 'ab(·)' in Eq. (2) is slightly inconsistent with 'ab(θl)' in Eq. (1).","section":"Notation, page 2"},{"comment":"The journal name contains a typo: 'IEEE Trans. V eh. Technol.' should be 'IEEE Trans. Veh. Technol.'","section":"Reference [3]"},{"comment":"Please clarify that the runtime comparison includes CA-CFAR latency for the signal-processing baselines but not for the learning-based methods, and state whether all methods run on the same GPU/CPU and use the same batch size.","section":"§IV-D"},{"comment":"The fine module uses a 256×256×256 3D FFT; the computational cost and memory footprint of this step are reported only indirectly through FLOPs. A brief discussion of memory consumption would help practitioners.","section":"§III-D"}],"recommendation":"major_revision","confidential_remarks":"The technical design is coherent and the paper addresses a timely problem, but the simulation-based evaluation is currently too opaque to support the headline claims. The most important issue is the unspecified RCS/scattering model in the Sionna RT dataset; without it, the cross-band fusion gain, which is concentrated in counting, could be a simulation artifact. I would also insist on multi-seed results and a corrected runtime claim. The paper is within the journal's scope and, after these revisions, could be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a new-method paper, not a new-physics paper. The architecture — count module, coarse transformer, fine patch refinement — is a sensible composition of known building blocks, and the paper is honest enough to show that the fine module does the heavy lifting for parameter accuracy. But the central claim, that learned cross-band fusion beats handcrafted fusion for counting, rests on the unstated RCS distribution in the Sionna RT simulation. If ρ_b,l are drawn independently per band, the 7.6-point count-accuracy gain over high-only is a simulator artifact, not a real multi-band effect. The stress-test concern is on target here.\n\nWhat the paper does well: the ablation is coherent, and the count-vs-SNR curve degrades gracefully outside the training range, which is a good sign. The design choices are explained — the velocity cross-attention, the gated fusion of heatmap and direct offsets — and the authors disclose in §IV-D that CSIYOLO-based is faster than their method, which suggests they are not hiding the embarrassing comparison.\n\nSoft spots, in order of severity:\n1. Per-band RCS is the load-bearing premise. §IV-A says only that the dataset was generated with Sionna RT; there is no description of how ρ_b,l or target RCS is set. That is the difference between a realistic digital twin and a synthetic toy. The authors need to specify the RCS model and, ideally, validate against measured dual-band RCS data.\n2. All headline numbers are single-run, with no seeds and no error bars. The 82.2%/56.9%/73.2% reductions are point estimates on one dataset.\n3. No code or dataset release. The dataset is private, so the numbers cannot be independently reproduced.\n4. The runtime claim in the abstract is true but only against conventional dual-band pipelines; the fastest learning baseline is 2.4× faster. The paper does disclose this later, but the abstract should be more careful.\n\nNone of these are deal-breakers if the authors treat them as conditions. The central argument — that a learned cascade with cross-band attention can outperform handcrafted fusion in a specific simulated campus scenario — holds up as a simulation result. The paper does not claim real-world validation.\n\nWho this is for: people working on learning-based ISAC sensing and multi-band fusion, particularly for drone detection in urban settings. It is worth a serious referee, mainly to push the authors on the RCS model and reproducibility. I would not cite it as a benchmark until code and RCS details are released, but it is a legitimate contribution to the simulation literature.\n\nRecommendation: send it to peer review. Require the authors to disclose the RCS simulation model, report multiple seeds or confidence intervals, and release code/data. If they cannot or will not, the paper's claims should be downgraded to “demonstration on a private simulation.”","headline":"DMSNet is a plausible learning-based fusion pipeline for dual-band ISAC, but its headline gains are only as good as the unstated per-band RCS simulation model, and the single-run numbers need variance and code.","tokens_in":10415,"tokens_out":3372,"would_cite":false,"duration_ms":30854,"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":"A neural network that learns to fuse 3.5 and 28 GHz echoes jointly counts targets and estimates their range, velocity, and angle, outperforming handcrafted dual-band ISAC methods in simulation.","keywords":["multi-band ISAC","cross-band fusion","target counting","multi-target parameter estimation","coarse-to-fine neural network","dual-band sensing","OFDM sensing","deep learning"],"falsifier":"Take DMSNet and the same baselines to a measured dual-band dataset, e.g., a 3.5 GHz and 28 GHz base station tracking real drones over an open field, and compare counting accuracy and median parameter errors. If DMSNet's counting-accuracy advantage over high-band-only processing shrinks to near zero, the cross-band complementarity premise fails. Alternatively, synthesize echoes with identical RCS across bands; DMSNet should then perform no better than a strong single-band network.","tokens_in":9181,"feed_emoji":"📡","tokens_out":4330,"duration_ms":40804,"temperature":0.7,"pith_summary":"This paper proposes DMSNet, a neural network for dual-band integrated sensing and communication that jointly estimates the number of targets and their range, velocity, and angle. Existing dual-band methods rely on handcrafted fusion rules, mostly ignore target counting, and are computationally heavy. DMSNet encodes each band separately, fuses features with cross-band attention, and refines coarse estimates with local high-resolution patches. In a simulated campus scenario with up to five drone targets, DMSNet raises exact counting accuracy to 91.74%, cuts median range, velocity, and angle errors by 82.2%, 56.9%, and 73.2% versus the best baselines, and runs in 23.7 ms per sample.","feed_headline":"Neural fusion beats handcrafted dual-band ISAC","feed_subtitle":"Count accuracy hits 91.7%; range, velocity, angle errors fall 82%, 57%, 73% in simulation.","key_machinery":"The central mechanism is multi-head cross-band attention over tokens extracted by 3D CNN encoders from each band's range-velocity-angle power spectrum, complemented by fusion features formed from the concatenation, absolute difference, and Hadamard product of the two bands' enhanced tokens. These fused tokens drive a counting head and, in the coarse module, serve as memory for a transformer decoder whose object queries are initialized from the counting output. The fine module then extracts 3D FFT patches around coarse centers and combines heatmap-based and direct-regression offsets through a learned gate.","core_discovery":"DMSNet treats high- and low-band echoes not as signals to be combined by fixed weights but as two views whose complementarity is learned. Its three cascaded modules—target counting, coarse grid classification with a transformer decoder, and fine residual regression on local dual-band patches—are trained with separate losses and executed sequentially. On a simulated dual-band dataset generated from a ray-traced campus digital twin at 3.5 and 28 GHz, the learned fusion outperforms the strongest counting baseline by 2.7 points in count accuracy and by 2.3 points in Macro-F1, and reduces median absolute errors of range, velocity, and angle by 82.2%, 56.9%, and 73.2% relative to the best conventi","pith_inferences":["If the simulated band-dependent scattering faithfully mirrors reality, the counting improvement indicates that learned fusion genuinely exploits targets visible in only one band; a real-world testbed with measured 3.5/28 GHz echoes would be the natural check.","The architecture generalizes in principle to more than two bands, since the attention fusion treats bands as tokens; whether gains grow or saturate as bands are added is untested here.","The 82.2% range-error reduction is largely an artifact of the fine module's local refinement; one could test whether a simpler interpolation around coarse bins achieves similar gains without learned offsets.","The method's reliance on per-module training (due to non-differentiable indexing) leaves open whether an end-to-end version would improve accuracy further; the authors do not report results from joint training."],"forward_implications":["Learned cross-band fusion can replace handcrafted weighting rules in multi-band ISAC sensing, removing the need to design band-combination strategies per scenario.","Target counting and continuous parameter estimation can be performed in a single forward pass, which existing conventional dual-band pipelines do not offer without a separate detection stage.","The coarse-to-fine local FFT refinement is the dominant contributor to parameter accuracy, suggesting that deep regression becomes unnecessary once spectral peaks are localized.","At 23.7 ms per inference, the network is fast enough for near-real-time multi-target sensing on current GPU hardware, while still outperforming conventional methods."],"fun_headline_variants":["DMSNet learns cross-band fusion for sharper ISAC sensing","Learned fusion lifts dual-band ISAC accuracy and speed","Neural network cuts ISAC errors by up to 82%","Cross-band learning beats handcrafted fusion in ISAC","DMSNet: 91.7% count accuracy, 68.7% faster"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The ray-traced synthetic dataset must reproduce realistic band-dependent target visibility and multipath; if real 3.5/28 GHz echoes do not show the same complementarity, the fusion gains—most of which appear in target counting—would not transfer from simulation.","fun_headline_variants_meta":{"raw":{"variants":["DMSNet learns cross-band fusion for sharper ISAC sensing","Learned fusion lifts dual-band ISAC accuracy and speed","Neural network cuts ISAC errors by up to 82%","Cross-band learning beats handcrafted fusion in ISAC","DMSNet: 91.7% count accuracy, 68.7% faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000637,"raw_usage":{"total_tokens":2762,"prompt_tokens":724,"completion_tokens":2038,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":1957}},"tokens_in":468,"tokens_out":2038,"duration_ms":13128,"temperature":1.0,"reasoning_tokens":1957,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T17:22:39.072453+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take DMSNet and the same baselines to a measured dual-band dataset, e.g., a 3.5 GHz and 28 GHz base station tracking real drones over an open field, and compare counting accuracy and median parameter errors. If DMSNet's counting-accuracy advantage over high-band-only processing shrinks to near zero, the cross-band complementarity premise fails. Alternatively, synthesize echoes with identical RCS across bands; DMSNet should then perform no better than a strong single-band network.","supporting_citations":[],"review_version":1}