{"id":"6cc6aae3-038e-4aae-b1e8-a73525bfa074","arxiv_id":"2607.25365","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A learned blockwise port-activation scheme for fluid antenna arrays lowers average peak sidelobes by 3.26 dB over uniform sparse activation at a slightly higher simulated sum rate, and by 8–10 dB over channel-driven selectors.","lead":"This paper trains a small neural network to choose which antenna ports a fluid antenna array should activate, balancing data rate against sidelobe radio leakage. In simulated 16-user downlinks, the learned selection lowers sidelobes by about 3–10 dB versus uniform, greedy, and gain-based rules while keeping throughput roughly equal.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline PSLL/rate gains are demonstrated only on the in-distribution synthetic channel model of Eq. 2; no out-of-distribution or measured-channel test supports the real-time FAA deployment claim.","rationale":"The reader's weakest assumption is the same: all evidence is in-distribution synthetic on Eq. 2, so the central claim lacks validation against channel-statistics mismatch. I do not see an internal arithmetic error: the reported 76.0% of full-port rate and the 3.26/8.13/10.10 dB PSLL deltas are consistent with Table I, and the paper is unusually transparent about the PSLL being a proxy and about full-port remaining the rate reference. The missing block partition and hand-set repulsion weights are real reproducibility gaps, but they are secondary: even if those were specified, the practical 'real-time FAA beamforming' claim would still depend on whether the learned activation transfers beyond the exact generator used for training and testing. A single out-of-distribution evaluation with a physically different FAA channel model would settle whether the learned scores and repulsion weights are genuinely channel-adaptive or merely tuned to this generator. Since the current manuscript does not provide such evidence, the conditional verdict is appropriate; my concern reinforces it rather than moving it.","tokens_in":13774,"tokens_out":8956,"duration_ms":98789,"concrete_test":"Take the trained L-BPA scorer and inference selector (with the missing block partition Bx×By and mb supplied by the authors) and evaluate it on a test set generated by the electromagnetic-aware FAA channel model of ref [22], or on measured port-domain channels from the same aperture geometry, with K=16, M=300 and the same RZF backend. Recompute Table I. If L-BPA's average PSLL advantage over Uniform drops below about 1 dB, or its average rate falls below Uniform, the headline claim is tied to the Eq. 2 generator and the real-time conclusion should be re-scoped. If the deltas persist, the external-validity concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Every reported number rests on a single synthetic channel law: Eq. 2 with UMi parameters, direction-cosine quantization, and per-user RMS normalization (Sec. IV-A). Training (1200), validation (200), and test (500) are all drawn from this law; the no-retraining sweeps in Fig. 7 only vary scalar parameters of the same law (SNR, angular spread, sector width). The channel model in Eq. 2 contains no mutual coupling, despite refs [22]–[24] showing that coupling materially reshapes FAA port-domain responses, and there is no near-field, dense-urban, or measured-channel test. Thus the abstract's 'real time ... FAA downlink beamforming' conclusion requires the unstated assumption that L-BPA's 3.26 dB PSLL gain and 0.58 bit/s/Hz rate advantage over Uniform survive channel-statistics mismatch. The paper honestly discloses that PSLL is a geometry proxy (Sec. II-C) and that full-port remains superior, but it does not disclose any sensitivity of the headline numbers to the channel law itself. This is an external-validity gap, not an internal inconsistency.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes L-BPA, a learned blockwise port activation method for fluid antenna array (FAA) downlink beamforming. A lightweight convolutional scorer maps multiuser channel features, port coordinates, and power statistics to port logits; blockwise hard masks are trained with a straight-through Gumbel-softmax relaxation and a differentiable peak sidelobe level (PSLL) surrogate. At inference, learned scores are combined with a multiscale geometric repulsion rule, and the resulting sparse port set is used for regularized zero-forcing (RZF) precoding. In simulations on a synthetic UMi far-field channel model with N=1250 candidate ports, K=16 users, and M=300 active ports, L-BPA achieves 98.06 bit/s/Hz average sum rate and -15.64 dB average PSLL, versus 97.48 bit/s/Hz and -12.38 dB for Uniform, 87.60 bit/s/Hz and -7.51 dB for a Greedy baseline, and lower PSLL for gain-based selection. The paper claims this shows real-time, sidelobe-aware, hardware-compatible port activation without iterative online search.","tokens_in":14046,"tokens_out":7316,"duration_ms":76262,"significance":"If the reported performance persists under realistic channel mismatch, L-BPA is practically attractive: it enforces grouped-switch constraints, avoids iterative RZF search, and explicitly trades sum rate against sidelobe behavior. The paper's strengths are its transparent problem formulation, explicit online-complexity analysis (Eq. 28), and ablations (Table II) that isolate the contributions of the learned scorer and the geometric loss. The principal weakness is that the entire evaluation rests on one synthetic channel law; moreover, the PSLL improvement is largely an explicit training objective, and the Greedy baseline is prefiltered. These issues do not invalidate the method but materially limit the strength of the claims as written.","major_comments":[{"comment":"All training, validation, and test channels are drawn from a single synthetic far-field sum-of-rays generator (Eq. 2 with UMi parameters). The no-retraining sweeps in Fig. 7 vary only scalar parameters of that same generator. No measured channels, mutual-coupling model (see refs. [22]–[24]), near-field, or different-geometry test is reported, despite the abstract's 'real time ... FAA downlink beamforming' deployment claim. For a data-driven method this is a load-bearing external-validity gap. Please add at least one out-of-distribution or measured-channel experiment, or substantially restrict the claims to the synthetic scenario.","section":"Sec. IV-A / Eq. (2) / Fig. 7"},{"comment":"The Greedy baseline is restricted to a candidate pool 'prefiltered by aggregate channel power' before sequential RZF-based selection. The headline comparisons against Greedy (8.13 dB PSLL improvement, 10.46 bit/s/Hz rate improvement) are therefore against a handicapped baseline unless the same prefilter is applied to L-BPA or the pool is large enough to be inconsequential. Report the prefilter size and provide a comparison to an unfiltered greedy (possibly with a complexity cap) so that the comparison actually tests the proposed method against the advertised greedy algorithm.","section":"Sec. IV-A / Table I"},{"comment":"The training loss explicitly includes the PSLL surrogate L_psll with target Γ_psll=-15 dB, and the reported average PSLL is -15.64 dB. Thus the '3.26 dB improvement over Uniform' is substantially a trained-for consequence of the objective rather than an emergent property. This is not a flaw per se, but the abstract's phrasing 'L-BPA reduces the average PSLL by 3.26 dB' overstates the finding. Please include a sweep of Γ_psll (or λ_psll) to show the achievable rate–PSLL tradeoff and clarify that the reported operating point is selected by the designer, not an unconstrained result.","section":"Eq. (24) / Table I"},{"comment":"The PSLL metric is the equal-amplitude broadside array factor (Eq. 10), not the actual RZF-weighted multiuser beam pattern. Since the manuscript's title and abstract concern beamforming, the sidelobe claims should be validated on at least a few representative RZF-weighted beam patterns, or the proxy nature of the metric should be made prominent in the abstract and conclusions. As written, the 3.26 dB improvement may not translate directly to the radiated multiuser beams.","section":"Sec. II-C / Figs. 3–4"}],"minor_comments":[{"comment":"Typos: 'termedgeometry' and 'areal timeandhardware' are missing spaces.","section":"Sec. I"},{"comment":"The normalization of z to \\bar z is not defined. Specify whether it is z-scoring, min-max, or another normalization.","section":"Algorithm 1, step 11"},{"comment":"The Gumbel-softmax relaxation uses clip(m_b softmax(...),0,1); after clipping the sum may not equal m_b. Clarify whether the exact cardinality constraint is enforced in the soft mask or only in the hard mask.","section":"Eq. (20)"},{"comment":"Report confidence intervals or paired significance tests for the 0.58 bit/s/Hz rate difference between L-BPA and Uniform; with 500 samples this difference could be within noise.","section":"Table I"},{"comment":"References [8] and [17] appear to be the same paper; please merge or differentiate.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the simulations are internally consistent, but the external-validity gap and the prefiltered Greedy baseline are substantial enough to require revision before publication. The missing experiments (out-of-distribution channels, unfiltered greedy, Γ_psll tradeoff) are feasible and should be obtained. I do not recommend rejection, but the current abstract overclaims relative to the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: L-BPA is a sensible engineering combo—CNN port scoring, straight-through Gumbel masks, a differentiable PSLL surrogate, and a blockwise repulsion rule—that answers a real need in FAA downlink: choosing a hardware-feasible binary mask without iterative RZF search. The paper is honest and the internal arithmetic holds up. What is new is the specific blockwise-hardware-constrained pipeline with multiscale geometry repulsion; the components are standard but the integration is new relative to the cited prior work. The ablations show all parts matter.\n\nWhere it gets shaky: the headline PSLL gain is partly circular. The training loss explicitly minimizes a PSLL surrogate with target Γ=-15 dB, and the reported average PSLL is -15.64 dB. So the 3.26 dB improvement over Uniform is a check of the optimizer, not an emergent discovery. That is not disqualifying—if you train for sidelobe suppression you should get it—but the paper should be read that way.\n\nThe bigger issue is external validity. Every number comes from one synthetic UMi channel generator (Eq. 2), with train, validation, and test drawn from the same law, and the no-retraining sweeps only vary scalars of that same generator. No mutual coupling, no near-field, no dense-urban, no measured channels—despite the authors' own refs [22]–[24] showing coupling changes FAA port responses. So the 'real time FAA beamforming' claim is a deployment claim the evidence doesn't support yet. Also missing: code/data, and the block partition (B_x × B_y and m_b) is never specified, so the complexity figures in Eq. (28) are not fully checkable. The Greedy baseline is handicapped by a high-gain prefilter, which is disclosed but makes the 8.13 dB comparison flattering.\n\nNone of this is a fatal flaw; the paper is re-implementable from the text and the math is clean. I'd send it to review, but I'd ask for an out-of-distribution or measured-channel test, code release, and a statement of the block partition. Without those, it's a well-formed idea with a narrow evidence base.\n\nWho it's for: FAA and 6G people working on port selection; they'll find it useful even if they don't buy the headline. I'd take it to reading group and cite the method, not the numbers.","headline":"A sensible, internally consistent combination of standard learning and geometry tools for FAA port selection, but the headline PSLL gain is largely trained-for and all numbers rest on one synthetic channel model; should go to peer review with a request for external-validity tests and code/data.","tokens_in":14653,"tokens_out":2544,"would_cite":true,"duration_ms":27376,"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 blockwise port selector can cut fluid-antenna sidelobes by 3.26 dB while keeping sum rate.","keywords":["fluid antenna array","port activation","sparse array beamforming","sidelobe suppression","regularized zero forcing","neural network port selection","blockwise activation","multiuser downlink"],"falsifier":"Evaluate L-BPA on measured or full-wave-simulated fluid-antenna channels that include mutual coupling and near-field effects; if the PSLL gap to uniform sparse activation shrinks below a few decibels or average sum rate drops below Uniform, the paper's central claim is falsified.","tokens_in":13544,"feed_emoji":"📡","tokens_out":3360,"duration_ms":33387,"temperature":0.7,"pith_summary":"The paper is trying to establish that the hard combinatorial problem of choosing which ports of a fluid antenna array to switch on can be solved in real time by a small neural scorer plus a deterministic geometry rule, without iterative search. It claims this learned blockwise port activation (L-BPA) outperforms uniform sparse activation by 3.26 dB in average peak sidelobe level while slightly increasing sum rate, and beats greedy and gain-based selection by 8.13 dB and 10.10 dB in PSLL. The practical stakes: FAA downlinks could get channel-adaptive, sidelobe-aware port selection at the channel time scale, compatible with grouped-switch hardware.","feed_headline":"Learned port selector cuts fluid-antenna sidelobes 3.26 dB","feed_subtitle":"Blockwise activation keeps sum rate up and needs no iterative search, enabling real-time FAA beamforming.","key_machinery":"Blockwise activation structure (each aperture block contributes a fixed number of ports, enabling grouped-switch hardware and preventing clustering); a small Conv–GN–SiLU scorer producing channel-aware port logits; block-wise straight-through masks that keep the discrete budget during training; a differentiable log-sum-exp PSLL surrogate that gives gradient signal for sidelobes; and a multi-scale geometric repulsion rule at inference that subtracts a proximity penalty from learned scores before selecting top ports per block. Together they convert an NP-hard subset-selection problem into one CNN forward pass plus O(sum(N_b^2 + m_b N_b)) deterministic selection.","core_discovery":"The central claim is that the rate–PSLL tradeoff in FAA downlink beamforming can be handled by decoupling the role of learning from the role of feasible selection. A lightweight convolutional network scores every candidate port from multiuser channel features, port coordinates, and power statistics; a block-wise constraint fixes the number of active ports per aperture block; a multi-scale repulsion term prevents local clustering; and the final mask feeds a regularized zero-forcing precoder. Trained end-to-end with straight-through masks and a differentiable peak sidelobe surrogate, the system reaches 98.06 bit/s/Hz average sum rate (about 76% of the full-port reference) at −15.64 dB average","pith_inferences":["If the gains transfer beyond the paper's synthetic far-field UMi channels, the same blockwise architecture could serve as a generic low-latency port-selection front end for any reconfigurable-aperture array, since the geometry rule is channel-agnostic.","The fixed repulsion parameters and block size are tuned for a 5λ×2.5λ aperture; a natural extension would be to learn them per aperture or per channel statistics, which the paper does not explore.","The paper's future-work mention of wideband beam-squint suggests that extending the PSLL surrogate to frequency-dependent patterns is a direct next test; one could check whether the 3.26 dB gain persists across subcarriers."],"forward_implications":["With only 24% of ports active, L-BPA keeps 76% of full-aperture sum rate while improving PSLL by 3.26 dB over a uniform sparse grid, so sparse FAA operation need not sacrifice sidelobe quality.","Because inference requires no iterative RZF re-evaluation over candidate masks, port activation can be recomputed at the channel time scale, compatible with real-time downlink scheduling.","The blockwise constraint matches grouped-switch hardware, meaning the method's mask is feasible to implement without per-port switches.","The ablation shows that removing either the learned scorer or the geometric PSLL surrogate degrades the PSLL/rate balance, indicating each component contributes."],"fun_headline_variants":["AI port selection trims fluid-antenna sidelobes 3.26 dB","Learned block port selection: −3.26 dB sidelobes, real-time","No-search FAA port learning: sidelobes down 3.26 dB, rate up","Neural port scores cut sidelobes 3.26 dB, keep sum rate"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central numbers come only from channels drawn by one synthetic far-field generator; if real channels (with mutual coupling, near-field effects, or other geometry) behave differently, the reported sidelobe and rate gains may not survive.","fun_headline_variants_meta":{"raw":{"variants":["AI port selection trims fluid-antenna sidelobes 3.26 dB","Learned block port selection: −3.26 dB sidelobes, real-time","No-search FAA port learning: sidelobes down 3.26 dB, rate up","Neural port scores cut sidelobes 3.26 dB, keep sum rate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000929,"raw_usage":{"total_tokens":3829,"prompt_tokens":773,"completion_tokens":3056,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":2976}},"tokens_in":517,"tokens_out":3056,"duration_ms":20498,"temperature":1.0,"reasoning_tokens":2976,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T02:38:33.904764+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate L-BPA on measured or full-wave-simulated fluid-antenna channels that include mutual coupling and near-field effects; if the PSLL gap to uniform sparse activation shrinks below a few decibels or average sum rate drops below Uniform, the paper's central claim is falsified.","supporting_citations":[],"review_version":1}