{"id":"15e5660b-efca-4cf8-a1b0-1610a1383d1e","arxiv_id":"2608.05896","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GSBF renders a 3D Gaussian scene into an angular propagation map and converts it into constant-modulus beamforming vectors, achieving higher spectral efficiency than codebook beam sweeping in simulation.","lead":"This paper trains a 3D Gaussian model of a room to generate antenna beam patterns directly from the access point's location and the user's position, skipping the usual step of measuring the wireless channel. If it works, it could cut the pilot overhead and beam-sweeping delay in millimeter-wave networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The deepest load-bearing risk is the first-order scattering rasterizer: Eqs. (10)-(13) model each Gaussian as an isolated secondary source, so multi-bounce paths have no representable forward path and cannot be absorbed by kernel learning.","rationale":"The reader identified the single-scattering approximation as the weakest assumption, and I agree. This is the most load-bearing concern because it is a structural restriction of the forward model: the rasterizer in Eqs. (10)-(13) enumerates paths of the form user -> primitive -> AP only. Even a very expressive Bi-SG kernel attached to a single primitive cannot emulate two distinct reflection points, because the outgoing ray from the first reflection must point toward the second reflector, not toward the AP, and the rasterizer provides no second-scattering step. Thus the claim that GSBF approaches the digital beamforming upper bound in real or ray-traced environments with multiple-bounce propagation is not established by the reported single synthetic scene. The phase-representation issue is related but secondary: the rasterizer also omits explicit path-length phase, which limits coherent multipath combining, but the learned kernels may partially compensate via narrow angular lobes; the single-bounce structural limit, however, has no analogous compensation. The reader's CONDITIONAL verdict already reflects this uncertainty, so I recommend no change to the verdict rather than a more severe penalty. The concrete test of varying Sionna's reflection order would directly determine whether the concern lands, because it isolates the effect of multi-bounce paths on GSBF's beamforming accuracy.","tokens_in":9161,"tokens_out":16306,"duration_ms":166770,"concrete_test":"Retrain the GSBF model on a scene with a deliberately placed pair of perpendicular metal plates forming a corner reflector, with the direct LoS blocked, and run Sionna twice: once with maximum reflection order = 1 and once with order = 3. Train separate models on the order-1 and order-3 datasets, then evaluate each on both test sets. If the model trained on order-1 data loses more than, say, 1 bps/Hz mean SE on the order-3 test set (or fails to steer toward the double-bounce direction at user positions where only the two-bounce path is strong), the first-order surrogate is insufficient. A secondary check: train on order-3 data and inspect whether the learned primitive placement can reproduce the double-bounce path; if it cannot, the structural limitation is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that GSBF can replace per-slot CSI depends on the rasterizer in Section IV-D faithfully approximating the Sionna ray tracer (or real propagation). The strongest structural limitation is that the renderer computes only first-order scattering: in Eq. (11), each primitive contributes once as o_i * tau_UE_i * tau_AP_i * S_i(omega_in, omega_out) * G', with tau_UE and tau_AP being multiplicative occlusion factors. There is no mechanism for the signal to scatter off one Gaussian and then a second, so a path that requires two reflections (e.g., a corner-reflector double bounce) has no representable form in the forward model. The learnable Bi-SG parameters cannot absorb such paths because they parameterize the angular response of a single primitive, not a sequence of primitives. If the real or simulated environment has significant higher-order reflections, the geometry-only inference stage will systematically miss those angular components, and the gap to DBF will grow well beyond the constant-modulus gap. The paper does not report experiments that vary reflection order, nor does it provide error bars or out-of-scene generalization, so the consistency of the reported 0.5-0.7 bps/Hz margin over EBA in Table I cannot be attributed to the physical model rather than to the particular Sionna configuration.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GSBF, a CSI-free beamforming framework that renders an angular propagator map from a persistent 3D Gaussian representation of the environment. The environment is represented by Gaussian primitives augmented with learnable bidirectional spherical Gaussian (Bi-SG) scattering kernels. A panoramic equirectangular projection and a two-sided electromagnetic rasterizer translate the AP-user geometry into a complex-valued angular propagator map, which is then aggregated through an over-complete array-manifold dictionary and projected onto the constant-modulus set to produce an analog beamforming vector. Training optimizes the Gaussian and kernel parameters on simulated CSI; at inference, only the AP pose and user position are needed. The paper reports simulations in a Sionna factory scene showing that GSBF outperforms exhaustive beam alignment (EBA) and a LoS-only geometric baseline (LocBF) while approaching the digital beamforming upper bound.","tokens_in":9432,"tokens_out":10324,"duration_ms":94070,"significance":"If the approach is validated, GSBF could provide a new way to reduce pilot and beam-sweeping overhead in mmWave MIMO systems by replacing per-slot CSI acquisition with a learned, geometry-consistent environment representation. The paper has several strengths: the problem formulation is clear, the differentiable rasterization pipeline is technically coherent, the use of an over-complete dictionary to synthesize constant-modulus beams is sensible, and the simulation results show a consistent margin over EBA. These strengths are offset, however, by important gaps in the physical model and in the evaluation, as detailed below.","major_comments":[{"comment":"Equation (11) models each primitive's contribution as a single scattering event along the path user → primitive i → AP, and there is no term for the direct user-AP path. The transmittances in Eqs. (10) and (12) are multiplicative occlusions, not secondary scattering events, so the rasterizer has no mechanism for the LoS component or for a path that reflects off two or more primitives. Since the ECDF discussion in Section V-B states that a subset of test positions has LoS paths, the missing direct-path term means those positions cannot be correctly represented by the rendered propagator map. Similarly, any environment with significant double-bounce propagation has no representable forward path, and the per-primitive Bi-SG parameters cannot absorb such sequences because they parameterize a single primitive's angular response, not a sequence of primitives. The authors should add an explicit LoS contribution and either extend the rasterizer to higher-order scattering or report a reflection-order analysis for the Sionna scene that demonstrates these missing terms are negligible.","section":"Section IV-D, Eq. (11)"},{"comment":"The only baselines are EBA, a DFT codebook search, and LocBF, a LoS-only geometric beamformer. No comparison is made to learning-based CSI-free methods such as a position-to-beam neural network, CKM-based beam alignment [4], or the DL location-based beamforming of [6]. Without such comparisons, the improved performance over EBA does not establish an advantage over the state of the art in environment-aware beamforming. Moreover, only mean and median SE over a single train/test split are reported, without standard deviations, confidence intervals, or multiple seeds, so the claim that GSBF 'consistently outperforms' EBA is not statistically supported.","section":"Section V-B, Table I"},{"comment":"The contribution of the physically motivated components (panoramic projection, Bi-SG kernel, two-sided rasterization) is not isolated. Because the Gaussian parameters and kernel coefficients are trained end-to-end on simulated CSI, a strong baseline such as an MLP that maps (AP pose, user position) directly to a beamforming vector would test whether the explicit environment representation adds predictive power beyond a learned black-box mapping. No such ablation is reported. This is important because the paper's central claim is that the environment representation is what enables CSI-free beam synthesis.","section":"Section IV-E"},{"comment":"The evaluation is confined to a single synthetic factory scene, and training and testing use channels generated by the same Sionna configuration. There is no test on a different scene, on the same scene with altered furniture or object layouts, or under different material parameters. Since the learned Gaussian representation is site-specific and static, the paper should at least characterize how performance degrades when the environment changes; otherwise the 'environment-aware' claim is only validated for one fixed configuration.","section":"Section V-A"}],"minor_comments":[{"comment":"The paper does not report training hyperparameters such as the optimizer, learning rate, number of iterations, or the number of Gaussian primitives in the factory scene, which hampers reproducibility.","section":"Section V-A"},{"comment":"The latency values are given without specifying the hardware platform and without comparing to the inference time of LocBF; please state the hardware and measure all baselines under the same conditions.","section":"Section V-B"},{"comment":"The definitions of u and v use θ and ϕ with arctan2, but the text should explicitly state the range conventions for the azimuth and elevation angles to avoid ambiguity.","section":"Section IV-C"},{"comment":"There are minor formatting issues, including missing spaces in 'withN x' and the expression 'W ′ = 1 2 W'; these should be corrected.","section":"Section IV-D"},{"comment":"The hyperparameters ν and ξ in the blending weight are set to 80 and 0.95 without any sensitivity analysis; a brief comment on their effect on performance would be useful.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"The central idea is interesting and within the scope of the journal, but the physical model has a significant gap (no LoS path and no multi-bounce mechanism), and the experimental validation is weaker than the claims require. The paper also cites several 2026 preprints, including the authors' own companion work [10], which are not yet peer-reviewed; the editor may wish to verify availability. No code or data is provided, which limits reproducibility. With the requested extensions (LoS term, additional baselines, error bars, and cross-scene evaluation), the paper could become a solid contribution, but in its current form the load-bearing claims are not fully supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing you should know: this paper is a genuinely new combination — 3D Gaussian splatting used to render a complex-valued angular propagator map that directly synthesizes analog beamforming vectors. Prior 3DGS work for wireless rendered real-valued power maps, so the complex rendering target and the reciprocity-preserving Bi-SG kernel are real contributions. The architecture is clearly described and the equations hang together.\n\nWhat it does well: the pipeline is coherent. The panoramic projection Jacobian is standard, the two-sided transmittance computation is explicit, and the final constant-modulus projection with straight-through estimation is a sensible way to train. The synthetic results show GSBF beating EBA by about 0.5-0.7 bps/Hz across array sizes with much lower latency, and the ECDFs give some insight into the LoS/NLoS mix. That is plausible.\n\nThe soft spot: the stress-test concern is right, and it is load-bearing. Eq. (11) defines each primitive's contribution as its own kernel times two transmittance factors; there is no term for a ray that scatters off one primitive and then a second before reaching the AP. So double-bounce and higher-order paths are not representable in the forward model. The Bi-SG parameters cannot learn their way around this, because they describe single-primitive scattering, not sequences. If the target environment has strong double-bounce reflections, the geometry-only inference will miss those angular components. The paper does not quantify this; it never varies the reflection order in Sionna or checks how much of the channel energy is multi-bounce.\n\nThere are also validation gaps: no error bars beyond mean/median (though the ECDF helps), no comparison against any learned beamforming baseline, only a weak LoS baseline, and no code or data release, so independent reproduction is impossible. The results are for one synthetic factory scene; the paper does not discuss generalization to other environments. None of these are fatal if the paper is positioned as a proof of concept, but they narrow what we can conclude.\n\nBottom line: this is a serious paper with a real idea. It deserves a serious referee — the novelty is there and the writing is honest about the architecture. The referee should push for either a model with explicit higher-order scattering or a careful empirical argument that first-order bouncing dominates in the target scenarios, plus stronger baselines and error bars. I would take it to our reading group.","headline":"Novel 3DGS-based beamforming with a real idea, but the first-order scattering rasterizer is a load-bearing limitation and validation is thin.","tokens_in":10012,"tokens_out":3275,"would_cite":true,"duration_ms":30528,"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":"A learned 3D scene of the environment can render analog beams directly, with no per-slot channel estimation.","keywords":["3D Gaussian splatting","beamforming","environment-aware communications","MIMO","channel state information-free","reciprocity","angular propagator map","multi-modal data"],"falsifier":"In a real indoor room whose measured channel is dominated by strong second- and higher-order reflections, compare the beamforming gain of a trained GSBF model against exhaustive beam alignment: if GSBF falls clearly below exhaustive search wherever multi-bounce paths dominate, the first-order scattering surrogate fails.","tokens_in":8945,"feed_emoji":"📡","tokens_out":5503,"duration_ms":49593,"temperature":0.7,"pith_summary":"This paper attempts to establish that a persistent 3D Gaussian representation of a wireless environment can replace per-slot channel state information for analog beamforming. The authors argue that radio propagation is governed by physical geometry, so once the scene is learned, the access point can synthesize a beam from its own pose and the user's position alone. They report that this approach approaches the digital beamforming upper bound while outperforming exhaustive codebook search and geometry-only line-of-sight beamforming in a simulated indoor factory. If true, the practical cost of beam training drops dramatically: beam synthesis becomes a fast, differentiable rendering step rather than a per-slot optimization problem.","feed_headline":"One learned 3D scene renders beams without per-slot CSI","feed_subtitle":"GSBF turns Gaussian splats into constant-modulus beams that beat codebook search and approach the digital upper bound.","key_machinery":"The load-bearing object is the Bi-SG kernel, which parameterizes each Gaussian primitive's scattering response $S_i(\\omega_{\\text{in}}, \\omega_{\\text{out}})$ as a complex albedo plus $K$ spherical-Gaussian lobes whose alignment metric blends collinear and redirected scattering, and which satisfies $S_i(\\omega_{\\text{in}}, \\omega_{\\text{out}}) = S_i(-\\omega_{\\text{out}}, -\\omega_{\\text{in}})$ for link reversal. This kernel sits inside a two-sided electromagnetic rasterizer: a source-side transmittance $\\tau^{\\text{UE}}_i$ accounts for shadowing between user and primitive, and an AP-side accumulation adds complex contributions $\\tau^{\\text{UE}}_i \\tau^{\\text{AP}}_i S_i G'$ into an angular propagator map. The map is vectorized, combined with an over-complete array-manifold dictionary $\\mathbf{A}$, and normalized to constant modulus, with straight-through estimation back-propagating through the projection. The machinery converts a static scene of learnable primitives into a differentiable, channel-relevant rendering that is decoupled from the array dimension.","core_discovery":"GSBF's central claim is that the environment's scattering response can be captured by Gaussian primitives augmented with reciprocity-preserving bidirectional spherical Gaussian (Bi-SG) kernels, and that a two-sided electromagnetic rasterization of these primitives—a source-to-primitive transmittance pass followed by a primitive-to-access-point coherent accumulation pass—renders a complex-valued angular propagator map. That map is combined with an over-complete array-manifold dictionary and projected onto constant-modulus entries to produce the beamformer. The authors claim this turns beamforming design into a learned forward map from geometry to beam, so the same trained scene synthesizes beams for arbitrary access-point and user geometry without instantaneous CSI, and they report that the resulting beams outperform exhaustive beam alignment and geometric baseline methods while staying close to the digital beamforming upper bound.","pith_inferences":["Not stated in the paper, but plausible: because the rasterizer decouples environmental scattering from the array response, the same learned scene could be re-rendered for a different array manifold or a second access-point viewpoint without retraining, turning it into a shared radio map for multiple basestations.","Not stated in the paper, but plausible: if occasional LiDAR or depth updates feed the density-control loop, the explicit Gaussian primitives might track slow environmental changes such as moving furniture or new obstacles without collecting fresh channel measurements.","Not stated in the paper, but plausible: the reciprocity symmetry of the kernel suggests the rendered angular propagator map could also predict uplink channels or initialize channel estimation, since the same environment response applies under link reversal."],"forward_implications":["Online inference needs only the access-point pose and user position, eliminating per-slot pilot CSI acquisition and iterative beamformer optimization.","Beam alignment latency stays roughly 21–24 ms as the array scales from 4×4 to 8×8, whereas exhaustive beam alignment's sweeping cost grows sharply with array size.","Because beams are synthesized over a dense angular grid rather than a fixed codebook, GSBF can steer more finely than codebook-based exhaustive search and reports higher spectral efficiency in the simulated indoor scene.","Spectral efficiency approaches the fully digital beamforming upper bound, with the remaining gap attributed to the constant-modulus hardware constraint and the learned representation's approximation error.","The persistent Gaussian scene is reused across user positions, so the expensive learning is amortized offline and only a fast rendering step is needed online."],"supporting_citations":[{"why":"Supplies the 3D Gaussian primitive representation, differentiable rasterization, and density control that GSBF builds on.","marker":"[11]"},{"why":"Provides the angular-domain superposition of steering vectors that justifies the over-complete dictionary synthesis of beamformers.","marker":"[15]"},{"why":"Supplies the spherical-Gaussian lobe parameterization used in the Bi-SG scattering kernel.","marker":"[16]"},{"why":"Supplies the bidirectional Gaussian parameterization that motivates the reciprocity-preserving incident/scattered response.","marker":"[17]"},{"why":"Provides the practical mmWave beam-sweeping timing used to compute exhaustive beam alignment's latency comparison.","marker":"[18]"}],"fun_headline_variants":["Gaussian splats beam without channel feedback","Beamform from 3D scene, no CSI needed","Scene-aware beams from splat rendering, no CSI","Splat-based beamforming beats codebook search, zero CSI","Render beams from geometry: GSBF ditches per-slot CSI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a scene made of independent, learnable scattering blobs, each bouncing the signal once, can faithfully imitate how real walls and objects scatter radio waves; if that imitation breaks down outside the training environment, the beams it predicts will point the wrong way.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian splats beam without channel feedback","Beamform from 3D scene, no CSI needed","Scene-aware beams from splat rendering, no CSI","Splat-based beamforming beats codebook search, zero CSI","Render beams from geometry: GSBF ditches per-slot CSI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1420,"prompt_tokens":889,"completion_tokens":531,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":450}},"tokens_in":505,"tokens_out":531,"duration_ms":5807,"temperature":1.0,"reasoning_tokens":450,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T21:31:03.250790+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a real indoor room whose measured channel is dominated by strong second- and higher-order reflections, compare the beamforming gain of a trained GSBF model against exhaustive beam alignment: if GSBF falls clearly below exhaustive search wherever multi-bounce paths dominate, the first-order scattering surrogate fails.","supporting_citations":[{"cited_title":"3D Gaussian splatting for real-time radiance field rendering,","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D Gaussian primitive representation, differentiable rasterization, and density control that GSBF builds on."},{"cited_title":"Efficient angle-domain processing for FDD-based cell-free massive MIMO systems,","cited_arxiv_id":null,"evidence_quote":"Provides the angular-domain superposition of steering vectors that justifies the over-complete dictionary synthesis of beamformers."},{"cited_title":"BiGS: Bidirectional Gaussian Primitives for Relightable 3D Gaussian Splatting","cited_arxiv_id":"2408.13370","evidence_quote":"Supplies the bidirectional Gaussian parameterization that motivates the reciprocity-preserving incident/scattered response."},{"cited_title":"A tutorial on beam management for 3GPP NR at mmWave frequencies,","cited_arxiv_id":null,"evidence_quote":"Provides the practical mmWave beam-sweeping timing used to compute exhaustive beam alignment's latency comparison."}],"review_version":1}