{"id":"79b3c9bc-4322-4967-a8bd-070cbf2ae6bb","arxiv_id":"2505.06277","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"RF-3DGS+ extends an existing 3D Gaussian splatting channel model to THz by embedding full propagation path length into the renderer, and claims accurate sparse-sampling reconstruction in simulation.","lead":"This paper tests whether a radio radiance field built with 3D Gaussian Splatting can reconstruct terahertz wireless channels from very few measured positions, and adds a fix that tracks the full path length of each reflected signal. It reports simulation results showing better reconstruction than three baselines in a synthetic indoor lobby, but the tests are limited to one simulated scene with no code or real measurements.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-bounce and pseudo-surface assumptions are load-bearing and unvalidated against multi-bounce ground truth; the claimed THz advantage is not established.","rationale":"The reader's weakest_assumption identifies the same load-bearing issue, and my reading agrees. The central claim that RF-3DGS+ captures key propagation paths with sparse training samples requires single-bounce paths to dominate the received power; otherwise Eq. 4 computes the wrong propagation distance, ToF, and interaction gain for exactly the paths that are claimed to be captured. The paper provides no multi-bounce ground truth and no error bound for the pseudo-surface approximation, so the empirical support in Section V is not independent of the assumption: if the Sionna simulation was configured with only single-bounce interactions, the comparison is circular. The proposed Sionna test directly resolves this by measuring multi-bounce energy and comparing reconstruction under both settings. No other concern is more decisive. The absence of released code and data weakens reproducibility but not the internal argument; the missing end-to-end communication metric is important but secondary to whether the reconstructed spatial spectra are correct in the first place. I therefore keep the reader's CONDITIONAL verdict: the method is plausible and internally consistent, but the decisive assumption needs direct validation. A positive multi-bounce test would support moving to ACCEPT; a negative one would require major revision or REJECT.","tokens_in":8667,"tokens_out":4494,"duration_ms":47682,"concrete_test":"In the same Sionna lobby, generate two ground-truth datasets: one with max_num_interactions=1, matching the implicit single-bounce setting, and one with max_num_interactions=5 to include double-bounce floor-wall and wall-ceiling paths. First compute the fraction of total received power carried by paths with two or more interactions in the multi-bounce dataset. Then train RF-3DGS+ identically on each dataset and compare PSNR/SSIM on the 100 held-out receivers, plus per-path AoD and ToF errors against the ray-traced ground truth. If the multi-bounce power fraction exceeds a few percent and RF-3DGS+ performance drops materially (e.g., PSNR decrease greater than 2 dB), the single-bounce dominance assumption is falsified for this environment; if the performance drop is negligible, the assumption is supported and the central concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV.B builds RF-3DGS+ on two assumptions: (1) in indoor THz, paths with two or more scattering reflections are negligible, so only single-bounce paths need be modeled; (2) the interaction point on each ray can be replaced by the center of the highest-contribution Gaussian (the pseudo-surface), which is used to compute lprev, lvd, and AoD in Eq. 4. Neither assumption is tested in Section V. The Sionna scene is configured with scattering parameters from [12] and [13], but the paper does not state the maximum number of bounces used to generate the ground-truth spectra, nor does it report the energy fraction carried by multi-bounce paths. If the simulator's ground truth excludes multi-bounce paths, the strong Table I results and the 20-sample resilience are guaranteed by construction. If multi-bounce paths are non-negligible, as they can be in a lobby with floor-wall or wall-ceiling combinations, Eq. 4's path length lprev+lvd and the single interaction gain product are wrong for those paths, and the claim that the reconstructed RRF captures key propagation paths is unsupported. The pseudo-surface approximation also has no error bound; a Gaussian center offset of even a few centimeters shifts ToF by more than the 0.4 cm multipath resolution quoted in Section III.A, so the 'minor errors' claim in Section IV.B requires quantification. Section VI explicitly admits the single-bounce limitation, making this the decisive, acknowledged gap in the argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RF-3DGS+, an extension of the RF-3DGS radio radiance field framework tailored to terahertz (THz) channels. The method uses a two-stage pipeline: first, scene geometry is reconstructed from visual data with 3D Gaussian splatting; second, a radio radiance field is trained from sparse RF measurements. The main novelty is the explicit incorporation of full propagation path information under a single-bounce scattering assumption: each Gaussian records a prior path length, AoD, and cumulative interaction gain, and the rendered path loss is computed from the total path length. The approach is evaluated in a synthetic Sionna indoor lobby scene against RF-3DGS, NeRF2, and CGAN, reporting higher PSNR/SSIM and lower LPIPS with much faster training and inference, plus resilience to very small training sets (down to roughly 20 samples).","tokens_in":9019,"tokens_out":4079,"duration_ms":41829,"significance":"If the central claims hold, this is a practically meaningful contribution: it shows that a radiance-field-style representation can capture THz spatial channels from sparse measurements, and the proposed path-length-aware rendering is a sensible remedy for the view-depth ambiguity that affects prior RF-3DGS. The paper is clearly written and the experimental setup, while synthetic, is grounded in Sionna with material parameters taken from THz measurement studies. However, the load-bearing single-bounce assumption and the pseudo-surface approximation are not validated against multi-bounce ground truth, and the reported quantitative comparisons lack variance and statistical detail. The evidence is therefore sufficient to motivate revision but not yet to establish the paper's strongest claims.","major_comments":[{"comment":"Section IV.B rests on the assertion that 'multi-bounce paths with two or more scattering reflections are usually negligible due to their high cumulative path loss,' but the paper never tests this assertion. The Sionna configuration in Section V reports EM parameters from Refs. [12,13] yet does not state the maximum number of bounces traced for the ground-truth spectra or the energy fraction carried by multi-bounce paths; if the simulator's ground truth excludes multi-bounce paths, the strong results in Table I and the 20-sample resilience in Fig. 4 are guaranteed by construction. Please add a multi-bounce ground-truth variant (or at least an energy-fraction report) and evaluate RF-3DGS+ against it, so that the paper's central claim is conditional on a testable assumption rather than on an input assumption.","section":"Section IV.B and Section V"},{"comment":"The pseudo-surface approximation, which replaces the interaction point on each ray with the center of the highest-contribution Gaussian, is claimed to introduce 'only minor errors,' but no error bound or ablation is provided. Because Section III.A quotes a multipath length resolution of 0.4 cm, a Gaussian-center offset of even a few centimeters changes the ToF by several resolution cells, so the adequacy of this approximation for THz bandwidths is not self-evident; please add an ablation that varies the pseudo-surface selection rule or compares against the true intersection point in the synthetic scene.","section":"Section IV.B"},{"comment":"The free-space path-loss factor (λ/(4π(lprev+lvd)))^2 is inserted into the renderer by construction, so the reported agreement on PSNR/SSIM in Table I and Fig. 4 largely tests the geometry from the visual stage and the learned directional gain, not the path-loss law itself. This does not invalidate the framework, but the paper should state explicitly which components are validated by the experiments and should report a diagnostic (e.g., error in estimated lprev+lvd against the simulator's true path length) to distinguish model assumption from learned quantity.","section":"Section IV.A, Eq. (4)"},{"comment":"Table I and Fig. 4 report point estimates without any measure of variance, number of runs, or statistical significance tests, and the min-max bands in Fig. 4 are not described at all; this makes the comparative claims about PSNR/SSIM and the sparsity threshold (Section V, 'as few as 20 samples') difficult to assess. Please report means and standard deviations over at least three independent training runs and describe the construction of the bands in Fig. 4.","section":"Table I and Fig. 4"}],"minor_comments":[{"comment":"The notation is inconsistent: the text refers to 'current view depth ldp' while Eq. (4) uses 'lvd.' Please unify the symbol for view depth.","section":"Section IV.A"},{"comment":"Eq. (4) uses N, Ai, AN, λ, lprev, lvd, and d_render without full definitions for all symbols; please add a notation table or define each term at first use.","section":"Eq. (4)"},{"comment":"Reporting wall-clock training time is not an apples-to-apples comparison across implementations; please specify common hardware and software versions and consider reporting per-epoch time or a normalized measure.","section":"Table I"},{"comment":"The caption of Fig. 4 does not describe the axes, the meaning of the shaded bands, or the number of trials; please add this information to make the figure interpretable.","section":"Fig. 4"},{"comment":"WRF-GS and Scalable 3DGS are discussed in Related Work as natural 3DGS-based baselines, yet neither is included in the experimental comparison; please state why they are omitted or include them.","section":"Related Work and Section V"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for eess.SP and the idea is a reasonable extension of RF-3DGS, but the central claims rest on an unvalidated single-bounce assumption and on simulation ground truth whose bounce configuration is not disclosed. Both points are fixable in revision, so I recommend major revision rather than rejection. No ethical or citation-pattern concerns beyond the noted missing discussion of baseline selection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick read of arXiv:2505.06277. It is a focused extension of the authors' own RF-3DGS to THz. The new piece is a full-path-length correction in the renderer: instead of sharing the same Spatial-CSI along a ray, they record prior path length lprev and add the view depth lvd when computing free-space path loss and ToF. Combined with a single-bounce assumption, this is a sensible way to remove the depth ambiguity that would otherwise hurt THz with its sub-centimeter multipath resolution. The simulation pipeline is coherent, and the comparisons to RF-3DGS, NeRF2, and CGAN are fair; the reported gains in PSNR/SSIM/LPIPS and training time are consistent with the design.\n\nThe soft spot is exactly what the stress-test note flags: the single-bounce assumption is load-bearing and unvalidated. Section IV.B asserts that multi-bounce paths are negligible due to cumulative loss, but the paper never reports the bounce order used in the Sionna ground-truth generation, nor the energy fraction carried by paths with two or more bounces. If the ground truth excludes multi-bounce, the strong results are guaranteed by construction. If it includes them, then Eq. 4's lprev+lvd and single-gain product are wrong for those paths. The pseudo-surface approximation (using the highest-contribution Gaussian center) also has no error bound; a few centimeters of offset shifts ToF beyond the quoted 0.4 cm resolution. These are not minor omissions; they are the two places where the claimed THz advantage either holds or collapses.\n\nThe rest of the evidence is thin: one synthetic lobby, a single table without variance or significance tests, and unspecified min-max methodology in Fig. 4. No code or data released. That said, the paper is honest about the single-bounce limit in the conclusion, and the idea is not silly. It is a plausible stepping stone for THz spatial channel modeling, not a proven one.\n\nMy recommendation: send it to peer review. The method is coherent and the problem is real; a referee can push the authors to validate the bounce assumption and release the Sionna configuration. I would not cite it yet, but I would not dismiss it either.","headline":"Focused extension of RF-3DGS to THz with a real depth-correction fix, but the single-bounce assumption is load-bearing and unvalidated.","tokens_in":9533,"tokens_out":1769,"would_cite":false,"duration_ms":16518,"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":"Radiance field maps THz channels from 20 sample points in simulation.","keywords":["terahertz communications","radio radiance field","3D Gaussian splatting","spatial channel modeling","spatial channel state information","single-bounce scattering","6G wireless","sparse channel reconstruction"],"falsifier":"In a ray-tracing scene that includes a large, smooth, highly reflective surface arranged so a transmitter-to-surface-A-to-surface-B-to-receiver path is comparable in energy to the single-bounce paths, compare RF-3DGS+ rendered path-loss and delay spectra to ground truth. If errors appear at the two-bounce path's angles and delays, the single-bounce dominance assumption is falsified; if the model renders accurately anyway, the assumption is supported.","tokens_in":8496,"feed_emoji":"📡","tokens_out":7514,"duration_ms":68488,"temperature":0.7,"pith_summary":"This paper argues that radio radiance fields, continuous learned maps of how radio energy leaves each surface point, remain workable in the terahertz band, where scattering rather than specular reflection is the main way signals reach non-line-of-sight receivers. It extends the 3D Gaussian splatting radiance framework RF-3DGS with explicit full-path information: for each emitting Gaussian, the model records the cumulative path length from the transmitter to the final scattering point and adds the receiver view depth, yielding unbiased free-space path loss and time-of-flight. Because THz paths beyond one bounce lose too much energy to matter, the model can reconstruct those paths from geometry alone and needs no delay or angle-of-arrival measurements as supervision. In a simulated indoor lobby, RF-3DGS+ reconstructs receiver-side spatial spectra with higher PSNR and SSIM, lower LPIPS, and far shorter training time than the compared baselines, and it keeps competitive quality down to about 20 training receiver positions. If this holds, sparse visual data plus a handful of RF measurements could supply the channel knowledge THz beamforming needs.","feed_headline":"Radiance field maps THz channels from 20 sample points","feed_subtitle":"In simulated indoor scenes, full single-bounce paths lift reconstruction quality and cut training to minutes.","key_machinery":"The mechanism is a radio radiance field rendered as $\\alpha$-blended 3D Gaussian primitives, where each Gaussian carries not only density and spherical-harmonic directional gains but also full propagation path metadata: cumulative prior path length $l_{\\mathrm{prev}}$, angle of departure at the transmitter, and the interaction gains accumulated up to the last bounce. During rendering, each Gaussian's path-loss contribution is $\\left(\\lambda/[4\\pi(l_{\\mathrm{prev}}+l_{\\mathrm{vd}})]\\right)^2$, so free-space loss and time of flight are computed from the true total distance instead of a shared per-ray value. The single-bounce assumption reduces path reconstruction to locating one interaction point per outgoing ray, and the model uses the center of the highest-contribution visible Gaussian as that pseudo-surface point. This object, a radiance field with explicit distance bookkeeping, is what carries the argument from geometry to accurate THz channel spectra.","core_discovery":"The central claim is that the ambiguity limiting radiance-field channel models at THz is a distance bookkeeping error, not a representation error, and correcting it is enough to make sparse-sample reconstruction work. In the predecessor model, all receivers along the same outgoing ray share one radiance value even though their distances from the final scattering point differ; at THz bandwidths the resulting path-loss and time-of-flight errors are large. RF-3DGS+ records for each Gaussian the prior path length $l_{\\mathrm{prev}}$ from transmitter to scattering point, the transmitter-side angle of departure, and the accumulated interaction gains, so the rendered path loss uses the true total distance $l_{\\mathrm{prev}}+l_{\\mathrm{vd}}$ and the delay is consistent with that distance. Under the single-bounce assumption the path is determined by the pseudo-surface point, approximated by the center of the highest-contribution Gaussian on the ray. The paper reports that this raises reconstruction PSNR from 13.50 to 19.68 and SSIM from 0.476 to 0.635 relative to the predecessor model in its simulated lobby, with inference at 3.4 ms per query.","pith_inferences":["The paper's own reliance on simulated data leaves open how the pseudo-surface approximation behaves with real surfaces that scatter from a patch rather than a point; a comparison against full ray tracing with extended scatterers would map where the approximation breaks.","Because the correction is just distance bookkeeping, the same idea could improve other radiance-field channel models, not only the Gaussian splatting family: any rendering equation that ignores view depth should gain accuracy in wideband settings.","A direct extension the authors do not run is a scene with a deliberately strong second-bounce path, such as a large metal panel, to locate the boundary of the single-bounce assumption; until that boundary is measured, the reported gains are a property of the single-bounce regime."],"forward_implications":["Sparse deployments become viable: with roughly 20 training receiver positions the reconstructed spatial spectra stay competitive, so a site could be mapped with a small measurement campaign rather than dense scanning.","Fast updates become possible: training finishes in minutes and inference in milliseconds, which is the regime needed for beam selection and handover decisions rather than offline planning.","Geometric priors substitute for expensive RF supervision: delay and angle-of-departure spectra are not needed as training labels because geometry plus the single-bounce assumption supplies them.","The approach inherits a scope condition: it is formulated for indoor and short-range THz links where single-bounce scattering dominates, and the authors explicitly point to multi-bounce modeling as future work."],"supporting_citations":[{"why":"Supplies the predecessor radio radiance field with 3D Gaussian splatting that RF-3DGS+ extends, and is the geometry-aware baseline whose view-depth ambiguity the paper corrects.","marker":"[11]"},{"why":"Provides the neural-radiance RF baseline that motivates the work and which RF-3DGS+ must beat in reconstruction quality and speed.","marker":"[9]"},{"why":"Provides the differentiable ray-tracing simulator used to generate the ground-truth THz spatial spectra and the simulated indoor scene.","marker":"[6]"},{"why":"Supplies measured transmission, reflection, and scattering characteristics of indoor building materials that set the surface EM parameters and motivate the scattering-dominated THz model.","marker":"[12]"},{"why":"Supplies measured diffuse-scattering behavior at THz frequencies, including the near-zero specular reflection coefficients that justify the single-bounce scattering focus.","marker":"[13]"}],"fun_headline_variants":["Radiance field fix makes THz sparse-sampling work","THz channel modeling: One fix lifts PSNR from 13.5 to 19.7","Single-bounce radiance fields cut THz training to minutes","Distance fix unlocks radiance field for THz channels","Corrected radiance field maps THz channels from sparse samples"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands or falls on single-bounce dominance: if two-bounce scattered paths carry meaningful energy in a real indoor THz scene, then the path-loss and time-of-flight corrections that produce the reported gains are missing those paths.","fun_headline_variants_meta":{"raw":{"variants":["Radiance field fix makes THz sparse-sampling work","THz channel modeling: One fix lifts PSNR from 13.5 to 19.7","Single-bounce radiance fields cut THz training to minutes","Distance fix unlocks radiance field for THz channels","Corrected radiance field maps THz channels from sparse samples"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000276,"raw_usage":{"total_tokens":1655,"prompt_tokens":962,"completion_tokens":693,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":601}},"tokens_in":578,"tokens_out":693,"duration_ms":6460,"temperature":1.0,"reasoning_tokens":601,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:41:36.077746+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a ray-tracing scene that includes a large, smooth, highly reflective surface arranged so a transmitter-to-surface-A-to-surface-B-to-receiver path is comparable in energy to the single-bounce paths, compare RF-3DGS+ rendered path-loss and delay spectra to ground truth. If errors appear at the two-bounce path's angles and delays, the single-bounce dominance assumption is falsified; if the model renders accurately anyway, the assumption is supported.","supporting_citations":[{"cited_title":"Sionna rt: Differentiable ray tracing for radio propagation modeling,","cited_arxiv_id":null,"evidence_quote":"Provides the differentiable ray-tracing simulator used to generate the ground-truth THz spatial spectra and the simulated indoor scene."},{"cited_title":"Transmission, reflection, and scattering characterization of building materials for indoor thz communications,","cited_arxiv_id":null,"evidence_quote":"Supplies measured transmission, reflection, and scattering characteristics of indoor building materials that set the surface EM parameters and motivate the scattering-dominated THz model."},{"cited_title":"A study of diffuse scattering in massive mimo channels at terahertz frequencies,","cited_arxiv_id":null,"evidence_quote":"Supplies measured diffuse-scattering behavior at THz frequencies, including the near-zero specular reflection coefficients that justify the single-bounce scattering focus."}],"review_version":1}