{"id":"1b966698-04ea-4f31-87a5-7550b15a3230","arxiv_id":"2501.10941","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A vertical federated learning scheme fuses heterogeneous vehicle sensors with short pilots to compute FDD precoders, reaching near-perfect-CSI WMMSE sum rates in simulation.","lead":"This paper proposes a vertical-federated-learning precoding method for FDD massive MIMO in which each vehicle computes its own precoding vector locally from a very short pilot plus its GPS, camera, and LiDAR data. Simulations in one city-block scenario show this method reaches about 92 percent of the sum rate of WMMSE with perfect channel knowledge while using far fewer pilot symbols.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (15) with λ=10 makes the training loss minimize, not maximize, sum rate; the central H-MVMM result cannot be reproduced with the stated loss and gradient descent.","rationale":"The reader's verdict is CONDITIONAL, and I agree that the paper should not be accepted without addressing key issues. However, I find a more immediate load-bearing concern than the reader's chosen weakest assumption. The reader's weakest assumption concerns the transferable value of sensing across scenarios and frequencies; they also mention 'a clear mathematical error in the loss-function explanation' in the rationale, but they do not make it the central condition for the verdict. My analysis shows that this error is not merely an explanation glitch: it reverses the training objective. With λ=10, Eq. (15) implies L = 9ΣR_k − 10K R_T, so gradient descent minimizes sum rate. This is internally inconsistent with the reported sum-rate results and with the claim that the additional term accelerates convergence. Even granting all the paper's assumptions about sensor informativeness, dataset suitability, and transferability, the stated algorithm should not produce the reported curves. This is a concrete, checkable issue that must be resolved. I therefore keep the conditional verdict but would make the condition explicit: the authors must correct Eq. (15) (or clarify the update direction) and confirm that the corrected training objective reproduces the reported numbers. If the sign error is simply a typo, the paper can be fixed; if not, the central claim collapses. This is not an ad hominem critique; it is a mathematical property of the written loss function.","tokens_in":10798,"tokens_out":6533,"duration_ms":80689,"concrete_test":"Re-run Algorithm 1 from a fixed initialization with the stated hyperparameters (K=7, SNR=30 dB, λ=10, R_T=0.3, ADAM lr=1e-4) on the M3SC training split, computing the loss exactly as in Eq. (15) and tracking validation sum rate. If the loss decreases while sum rate decreases, or training diverges, the published Table I/Fig. 6 results cannot come from the stated loss. The authors should then supply the corrected loss (e.g., L = −Σ_k (R_k − λ(R_T−R_k)) or a hinge penalty) and confirm the reported curves are reproduced with it. If the corrected sign reproduces the numbers, the paper needs only a typo fix; if not, the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim—that H-MVMM with L_P=8 approaches WMMSE w/ GT CSIT—depends entirely on the training procedure. As written, Eq. (15) defines L = −Σ_k (R_k + λ(R_T−R_k)). Substituting λ=10 gives L = 9Σ_k R_k − 10K R_T. Minimizing L with gradient descent therefore minimizes the sum rate, not maximizes it. Algorithm 1 explicitly uses gradient descent updates Θ ← Θ − η∇ΘL, so a correct implementation of Eq. (15) should drive sum rate downward, contradicting the reported performance in Table I and Fig. 6. The paper claims the extra term 'accelerates convergence' via the gap (R_T−R_k), but the sign is wrong whenever λ>1: larger R_k increases L and is penalized. This is an internal inconsistency, not a cross-scenario generalization issue—it affects the stated algorithm even on the M3SC dataset. Unless Eq. (15) is a typo, the update direction is actually gradient ascent, or the training procedure is misdescribed, the near-WMMSE claim is unsupported. The reader's weakest assumption about sensor transferability is secondary; the loss-function sign error is more load-bearing because it invalidates the quantitative results as reported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes H-MVMM, a vertical federated learning (VFL) framework for FDD massive MIMO precoding in which vehicles with heterogeneous on-board sensors (GPS, RGB cameras, LiDAR) and short pilot sequences (LP=8) compute local precoding vectors from their multi-modal data and received pilots, feed the quantized vectors back to the base station, and receive gradients for training. Three hand-crafted preprocessing pipelines convert raw sensing data into precoding-relevant features, and a customized loss and an online training strategy are introduced to handle data heterogeneity and changes in the number of users. Numerical results on the M3SC city-block scenario are reported to show that H-MVMM with limited feedback approaches ZF with perfect CSIT and, with sufficiently large feedback, approaches WMMSE with perfect CSIT while using far fewer pilots. The paper also compares VFL with centralized learning in terms of transmitted data volume and sum rate.","tokens_in":11116,"tokens_out":25112,"duration_ms":289173,"significance":"If the reported numerical results are reproducible, the paper addresses a timely problem: reducing pilot overhead and precoding complexity in FDD massive MIMO by exploiting multi-modal sensing. The use of a public dataset, strong perfect-CSI baselines (ZF, WMMSE), and an explicit VFL formulation are strengths, and the claimed reduction from N=128 pilot dimensions to LP=8 is substantial. However, the evidence is currently a single-scenario, single-run simulation with no error bars or released code. More importantly, the advertised customized-loss mechanism is not effective as written, and the handling of missing sensor modalities in the heterogeneous configuration is underspecified. The numerical results also do not clearly demonstrate an advantage of heterogeneity, since the homogeneous Pilot-LiDAR scheme outperforms the heterogeneous H-MVMM scheme at every reported K. These issues make the central claims defensible in principle but in need of substantial revision and additional verification.","major_comments":[{"comment":"The claimed customized loss does not behave as stated. Expanding Eq. (15) with lambda=10 gives L = 9*sum_k R_k - 10K*R_T, so minimizing L maximizes the sum rate; the sign is therefore not the problem. The actual problem is that the added term is affine in R_k: its derivative with respect to each R_k is lambda-1, independent of R_T-R_k. Consequently, the statement that a large gap (R_T-R_k) results in a large derivative is false, and the loss provides no gap-dependent weighting or acceleration of convergence. Please either replace Eq. (15) with a nonlinear gap-dependent loss and re-derive the gradient, or remove the acceleration claim and provide an ablation isolating any actual benefit of the proposed term.","section":"Sec. III-C, Eq. (15)"},{"comment":"The H-MVMM scheme is evaluated with vehicles having different sensor subsets (5/7 vehicles with GPS, 3/7 with RGB cameras, and 3/7 with LiDAR), but the paper does not specify how the local NN handles missing sensor modalities. Equations (13)-(14) and Fig. 5 present separate branches for GPS, RGB, LiDAR, and pilots, and the text assumes a vehicle is equipped with three sensors. There is no description of zero-padding, conditional branch selection, per-configuration architectures, or any other mechanism for variable sensor availability. This makes the training and inference procedure for the heterogeneous configuration underspecified and not reproducible.","section":"Sec. III-B and Sec. IV-A.2"},{"comment":"There is a training/inference mismatch in the feedback quantization. Section III-C states that more quantization bits can be used for uplink feedback during training, while Section IV-A.2 reports that all VFL-based schemes use B=2 quantization bits. Since the loss is evaluated on quantized precoding vectors through Eqs. (4b)-(4c), the paper must state the exact training-time bit budget and whether quantization is modeled in the forward pass. Without quantization-aware training or an explicit description, the reported 2-bit results may not reflect the actual system being evaluated.","section":"Sec. III-C and Sec. IV-A.2"},{"comment":"The central empirical claim rests on single runs in one M3SC scenario, with no error bars, random seeds, or code release. Moreover, Table I shows that VFL-Pilot-LiDAR outperforms H-MVMM at every reported K (e.g., 80.2 versus 75.2 bps/Hz at K=7), so the numerical results do not demonstrate that sensing heterogeneity improves sum rate; the near-WMMSE performance is driven mainly by the LiDAR modality. The abstract's unqualified statement that the approach closely approximates perfect-CSI methods also overstates the H-MVMM gap (75.2 versus 81.6 bps/Hz at K=7). Please report statistics over repeated training runs and qualify the claims according to the actual gaps and the sensor configurations used.","section":"Table I and Sec. IV-B"},{"comment":"The complexity and data-volume comparisons are incomplete. The claim that H-MVMM has complexity O(L_P N) omits the NN forward pass, YOLOv8 building detection, and LiDAR BEV compression, which dominate the per-inference cost; the phrase constant time complexity increase with respect to N and L_P is inaccurate for a linear-order term. Similarly, the data-volume formula W_FL = N_epoch * K * 0.5KB in Section IV-D counts only the uplink precoding feedback and ignores the downlink gradient transmission required by Algorithm 1 (line 7). These comparisons should be corrected or qualified, or the claims should be limited to the pilot-processing component.","section":"Sec. IV-B and Sec. IV-D"}],"minor_comments":[{"comment":"The text says normalized weight and height; the first should be normalized width, and the typeset formula for q_{j,i} contains a stray symbol that should be cleaned up.","section":"Sec. III-A2, Eq. (11)"},{"comment":"The online training strategy's use of downlink channel estimation for loss calculation should be reconciled with the earlier no-explicit-CSI premise; please clarify whether online adaptation requires channel estimates and quantify the additional pilot overhead this entails.","section":"Sec. IV-C"},{"comment":"Please define what w/ Feedback versus w/ 2-bit Quantization mean in Fig. 6(a), give the value of B used for the sufficiently large B case, and add ZF w/ GT CSIT results to Table I so the claim of surpassing ZF can be checked numerically.","section":"Fig. 6(a) and Sec. IV-B"},{"comment":"The high-frequency encoding in Eq. (8) is the NeRF positional encoding; please cite the original source (Mildenhall et al.) in addition to reference [17].","section":"Sec. III-A1, Eq. (8)"},{"comment":"The statement that quantization losses are minimal is not well supported by Table II, where VFL H-MVMM gives 75.2 bps/Hz versus 79.1 for centralized learning; please quantify and qualify this claim.","section":"Sec. IV-D"},{"comment":"The output dimension is written as R^{2N_t}, but the system model uses N for the number of antennas; please use consistent notation throughout.","section":"Sec. III-B, Eq. (14)"},{"comment":"There are several typos and notation inconsistencies, including coloumn (Eq. (1)), leverging (Sec. III-B), surronding (Sec. III-A3), the the in the funding footnote, and the use of both Q and B for feedback bits.","section":"Throughout"},{"comment":"The phrase reciprocal of the sum rate appears to mean the negative sum rate; please correct the terminology.","section":"Sec. III-C, after Eq. (15)"},{"comment":"Please specify activation functions, normalization layers, LiDAR voxel dimensions Lx, Ly, Lz, YOLOv8 training details, and the criterion for validation loss stabilization, as these are needed for reproducibility.","section":"Fig. 5 and Sec. IV-A.2"}],"recommendation":"major_revision","confidential_remarks":"The stress-test note's claim that Eq. (15) makes training minimize the sum rate is not correct: with lambda=10, L=9*sum R_k - 10K*R_T, so minimizing L maximizes the sum rate. The real issue is that the gap term has no effect on the gradient, so the advertised acceleration mechanism is vacuous. I recommend major revision rather than rejection because the core idea is plausible and the identified problems are correctable in a revised manuscript. The paper would also benefit substantially from code release and repeated-run statistics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this paper has a genuinely useful idea and some solid engineering, but its central numerical claim is currently unsupported because the loss function in Eq. (15) is written with the wrong sign for the stated training rule. I'd still send it to review, because the fix is probably a typo and the idea is worth referee time.\n\nWhat's actually new: the authors extend their prior beam-selection work to full FDD precoding using vertical federated learning with heterogeneous sensor configurations. The three preprocessing pipelines (GPS angle encoding, YOLOv8 building detection to multipath one-hot vectors, LiDAR BEV) are sensible and well-motivated. The online user-count adaptation strategy addresses a real practical problem. The complexity argument (O(L_P N) vs WMMSE) is clear and fair. The baselines—ZF/WMMSE with perfect CSI, and a BS-NN with full channel estimation—are appropriate.\n\nThe soft spots: first and foremost, Eq. (15). With λ=10, L = -Σ_k (R_k + 10(R_T - R_k)) = 9ΣR_k - 10K R_T. Algorithm 1 updates Θ ← Θ - η∇L, so minimizing L minimizes sum rate. That is the opposite of what Table I and Fig. 6 report. I believe this is a sign error—likely they meant L = -Σ_k (R_k - λ(R_T - R_k)) or a hinge-style penalty—but as written the algorithm cannot produce the reported results. The simulations need to be re-run or the typo explicitly confirmed.\n\nSecondary issues: no code or data release, no error bars, and only a single city-block scenario at one frequency. The claim that sensing features transfer across environments is plausible but untested. The paper would also be much stronger with a conventional short-pilot baseline (e.g., LS or MMSE channel estimation with L_P=8) instead of only perfect-CSI and full-CE references.\n\nBottom line: the architecture and problem formulation are worth reading, but the quantitative claims rest on a formula that contradicts the training procedure. This is for researchers working on sensing-aided communications and federated learning for MIMO. A serious referee could get this into shape, so I'd accept it for review, but I would not cite the numerical results until the loss function is corrected and the simulations verified.","headline":"Good idea, broken loss-function sign makes the central numbers unsupported as written.","tokens_in":11606,"tokens_out":2900,"would_cite":false,"duration_ms":32888,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A vertical federated learning loop using eight pilots and heterogeneous vehicle sensors matches or beats perfect-CSI precoding baselines.","keywords":["FDD precoding","multi-modal sensing","vertical federated learning","massive MIMO","pilot overhead reduction","sensor heterogeneity","sum rate optimization","vehicular networks"],"falsifier":"Train the identical H-MVMM scheme on a different city-block scenario or at a different carrier frequency and check whether the sum-rate gap to WMMSE with perfect CSI remains small at $L_P=8$; the claim fails if the scheme degrades to roughly pilot-only performance in the new setting.","tokens_in":10596,"feed_emoji":"📡","tokens_out":8847,"duration_ms":85401,"temperature":0.7,"pith_summary":"This paper tries to establish that a base station can design multi-user precoding vectors almost as well as if it had perfect downlink channel knowledge, using only eight pilot symbols and the sensor data already available on nearby vehicles. The proposed H-MVMM scheme runs inside a vertical federated learning framework: each vehicle's local neural network maps its own received pilots, GPS position, camera views, and LiDAR scans to its own precoding vector, and the base station only assembles the vectors and returns gradients. If the claim is right, FDD systems could shrink pilot overhead from a channel-sized message to a fixed short sequence and cut precoding complexity from cubic in the number of antennas to roughly linear. The supporting evidence is a simulation in one city-block scenario at 4.95 GHz, compared against zero-forcing and WMMSE baselines that assume perfect channel state information.","feed_headline":"Eight pilots plus vehicle sensors rival perfect-CSI precoding","feed_subtitle":"Distributed learning cuts pilots to eight symbols and reaches zero-forcing or WMMSE-level sum rates with near-linear complexity.","key_machinery":"The central mechanism is the H-MVMM scheme, a vertical federated learning stack that splits the sum-rate optimization across vehicles. It rests on three preprocessing transforms that convert raw sensing into precoding-relevant features: a high-frequency encoding of GPS-derived line-of-sight angles, a YOLOv8-based building detector that turns camera images into one-hot multipath-angle vectors plus an encoded vehicle heading, and a voxelized bird's-eye-view compression of LiDAR point clouds. These sensing features are concatenated with features extracted from the received pilot vector and integrated by a local network that outputs the precoding vector. A rate-gap term in the loss accelerates convergence for weaker users, and an online training strategy retrains the local branches when the user count changes, saving at least 90 epochs relative to training from scratch in the reported $K=5$ to $K=6$ transition.","core_discovery":"The central claim is that sensing heterogeneity can be turned into an advantage for FDD precoding instead of a nuisance. Under the vertical federated learning setup, each vehicle's local network maps the received pilot vector plus its available modality representations to a quantized precoding vector, and the base station reconstructs the precoding matrix and computes a sum-rate loss. The paper reports that H-MVMM with $L_P=8$ pilots approaches the sum rate of zero-forcing with ground-truth CSI when $K<5$ and surpasses it as the number of users grows, and that with sufficiently large feedback quantization it approximates the WMMSE benchmark with ground-truth CSI while beating a base-station neural network that receives full channel estimates. The complexity of the learned precoder is $O(L_P N)$, compared with $O(K^2N^3)$ for one WMMSE iteration. The per-modality results rank LiDAR ahead of RGB images and GPS in sum-rate contribution, with the H-MVMM combination retaining performance close to the best single modality.","pith_inferences":["Beyond the paper: the same pipeline should be retrained and tested in a second city or at a different carrier frequency before the near-perfect-CSI result is treated as general, since the sensing-to-channel link is the unproven transfer point.","Beyond the paper: comparing H-MVMM against compressed-sensing or other low-pilot channel estimators at the same $L_P=8$ would isolate how much of the gain comes from vehicle sensing rather than from end-to-end learned precoding.","Beyond the paper: the claim that sufficiently large feedback approaches WMMSE suggests sweeping the feedback bit budget $B$ and plotting the sum-rate curve, which would show where quantization ceases to be the bottleneck.","Beyond the paper: the online training result is shown for one user-count change at three SNRs; a sensitivity analysis over sensor configurations and user arrivals would define when the 90-epoch saving persists."],"forward_implications":["At $L_P=8$ pilots, the H-MVMM scheme approaches zero-forcing with perfect CSI for $K<5$ and exceeds it for larger $K$ at 30 dB SNR in the tested scenario.","With sufficiently many feedback bits, the scheme approximates WMMSE with perfect CSI at high $K$ and outperforms a base-station neural network given full channel estimates.","The time complexity of producing one precoding vector is $O(L_P N)$, effectively linear in antenna count and independent of $K$, against $O(K^2N^3)$ for one WMMSE iteration.","The vertical federated learning framework transmits only about 0.5 KB per vehicle per training epoch, versus hundreds of megabytes for centralized learning, with a small sum-rate penalty from quantization.","The online training strategy adapts the network from five to six users in at least 90 fewer epochs than offline retraining, enabling response to fluctuating user counts."],"supporting_citations":[{"why":"Supplies the M3SC city-block scenario, building layout, and vehicle sensing data used in all experiments.","marker":"[19]"},{"why":"Defines the zero-forcing baseline with ground-truth CSI that H-MVMM is claimed to approach and then surpass.","marker":"[5]"},{"why":"Defines the WMMSE baseline with ground-truth CSI and provides the $O(K^2N^3)$ per-iteration complexity comparison.","marker":"[6]"},{"why":"Provides the deep-learning FDD precoding approach that motivates bypassing explicit channel estimation and directly generating precoding vectors.","marker":"[11]"},{"why":"Introduces the Synesthesia of Machines concept that guides using multi-modal sensing in place of some radio pilots.","marker":"[16]"},{"why":"Supplies the high-frequency positional encoding used to transform GPS angles and vehicle orientation into network inputs.","marker":"[17]"},{"why":"Provides the YOLOv8 object detector used to locate buildings in camera images for multipath-angle extraction.","marker":"[18]"},{"why":"Demonstrates multi-modal sensing for beam selection that this paper extends from beam selection to full precoding design.","marker":"[15]"}],"fun_headline_variants":["VFL with mixed sensors cuts pilots to 8, rivals perfect-CSI","Heterogeneous sensing in VFL gives 8-pilot precoding near perfect CSI","Mixed sensors and short pilots via VFL match perfect-CSI precoding","VFL turns sensor variety into 8-pilot precoding with perfect-CSI sum rates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that a vehicle's GPS position, the buildings its cameras can see, and its LiDAR scan of the street are a reliable stand-in for the radio channel at 4.95 GHz in this scenario; if that sensing-to-channel link weakens in a new city or band, the near-perfect-CSI performance is not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["VFL with mixed sensors cuts pilots to 8, rivals perfect-CSI","Heterogeneous sensing in VFL gives 8-pilot precoding near perfect CSI","Mixed sensors and short pilots via VFL match perfect-CSI precoding","VFL turns sensor variety into 8-pilot precoding with perfect-CSI sum rates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000558,"raw_usage":{"total_tokens":2641,"prompt_tokens":917,"completion_tokens":1724,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":1637}},"tokens_in":533,"tokens_out":1724,"duration_ms":13805,"temperature":1.0,"reasoning_tokens":1637,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:48:40.897119+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical H-MVMM scheme on a different city-block scenario or at a different carrier frequency and check whether the sum-rate gap to WMMSE with perfect CSI remains small at $L_P=8$; the claim fails if the scheme degrades to roughly pilot-only performance in the new setting.","supporting_citations":[{"cited_title":"M 3SC: A Generic Dataset for Mixed Multi-Modal (MMM) Sensing and Communication Integration,","cited_arxiv_id":null,"evidence_quote":"Supplies the M3SC city-block scenario, building layout, and vehicle sensing data used in all experiments."},{"cited_title":"Linear transmit processing in MIMO communications systems,","cited_arxiv_id":null,"evidence_quote":"Defines the zero-forcing baseline with ground-truth CSI that H-MVMM is claimed to approach and then surpass."},{"cited_title":"Intelligent multi-modal sensing-communication inte- gration: Synesthesia of machines,","cited_arxiv_id":null,"evidence_quote":"Introduces the Synesthesia of Machines concept that guides using multi-modal sensing in place of some radio pilots."},{"cited_title":"Attention is all you need,","cited_arxiv_id":null,"evidence_quote":"Supplies the high-frequency positional encoding used to transform GPS angles and vehicle orientation into network inputs."},{"cited_title":"Real-time multi class helmet violation detection using few-shot data sampling technique and YOLOv8,","cited_arxiv_id":null,"evidence_quote":"Provides the YOLOv8 object detector used to locate buildings in camera images for multipath-angle extraction."},{"cited_title":"Integrated sensing and communications toward proactive beamforming in mmWave V2I via multi-modal feature fusion (MMFF),","cited_arxiv_id":null,"evidence_quote":"Demonstrates multi-modal sensing for beam selection that this paper extends from beam selection to full precoding design."}],"review_version":1}