{"id":"1422486d-de74-4715-a73c-0d8674235e37","arxiv_id":"2608.13249","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"IA-ZPA selects 64 of 256 fluid-antenna ports at 76.28 bit/s/Hz with -13.66 dB PSLL and 2.73 ms decision time, the best rate among methods meeting the -13.5 dB target.","lead":"A new port-activation method for fluid antenna arrays, IA-ZPA, combines a learned CSI scorer with a fixed checkerboard aperture rule and a light mutual-coupling penalty. In simulations it keeps 76.28 bit/s/Hz at a -13.66 dB sidelobe level while deciding masks in 2.73 ms, beating fixed masks and sidestepping greedy selection's 200 ms latency.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated proxy alignment: the learned scorer optimizes the RZF proxy Rsel, but the headline rate is the current-domain REM; nothing in the paper tests whether Rsel and REM order masks the same way, and the 0.11 bit/s/Hz coupling refinement cannot fix a mismatch.","rationale":"The reader's weakest assumption is the proxy alignment between Rsel and REM, and I share it. This is the single most load-bearing point because the method's training signal is Rsel, while the reported result is REM. The two rates differ in noise normalization, power scaling, and impedance effects. Table II shows the impedance kernel, the only current-domain-informed component of IA-ZPA, moves REM by only 0.11 bit/s/Hz, so the learned CSI scoring and checkerboard projection, both trained on Rsel, are responsible for the rest. The paper's own Table I shows greedy, which also uses Rsel, has higher REM (76.75) at a worse PSLL. Thus there is no evidence that the Rsel-optimal masks are REM-optimal among PSLL-satisfying masks. A concrete rank-correlation test would settle this. If Rsel aligns with REM, the paper is scientifically sound and should be conditional only on reproducibility; if not, the headline comparison is not robust and the central claim should be weakened. The paper deserves credit for a clean common evaluation protocol, honest module attribution, and explicit acknowledgment that terminated inactive ports and EEP-based fields are left out. The issue is not that the model is physics-free; it is that the learned part is optimized for a different objective than the one reported, and the paper does not establish alignment.","tokens_in":9185,"tokens_out":15334,"duration_ms":173658,"concrete_test":"On the same 500 test channels, compute Rsel (Eq. 2 at rho=10 dB) and REM (the full current-domain backend) for a diverse mask library: all masks in Table I, 1000 random masks satisfying the per-zone quota per channel, and zonewise perturbations of IA-ZPA masks. Report the Spearman rank correlation and the pairwise agreement probability of Rsel and REM. If the correlation is below about 0.7 or pairwise agreement is near chance, the central claim is not supported without retraining on a current-domain-aware surrogate (e.g., adding a differentiable gamma/voltage-aware term to Eq. 7); if the correlation is high, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparison is only trustworthy if the training-time objective Rsel(a) (Eq. 2) is a monotone stand-in for the evaluation-time constrained rate REM computed by Eqs. (11)-(14). Rsel is a unit-Frobenius RZF rate at an arbitrary reference SNR (10 dB plus training jitter) with no mutual coupling and no source-drive scaling; REM depends on the current-domain RZF metric Xi and on the scaling factor gamma, because the fixed receiver-noise term N0 in Eq. (15) does not scale with gamma. These are different functionals of the mask, so a mask that maximizes Rsel under the soft-PSLL loss need not maximize REM. Table II makes the concern concrete rather than resolving it: the mutual-impedance kernel changes REM by only 0.11 bit/s/Hz and gamma by 0.0004, so almost all of IA-ZPA's advantage comes from the CNN/checkerboard part, whose training signal is Rsel. Yet the paper never reports the correlation (per channel or pooled) between Rsel and REM, nor shows that a different mask satisfying the -13.5 dB mean-PSLL target cannot beat 76.28 bit/s/Hz. Because greedy, which also uses Rsel, already reaches 76.75 bit/s/Hz before PSLL enforcement, a PSLL-constrained greedy or a coupling-aware retrained scorer could plausibly outperform the reported number. The absence of released code/data and of error bars makes this ordering assumption the load-bearing point: if Rsel misranks masks relative to REM, the headline gain is an artifact of the heuristic proxy rather than a property of IA-ZPA.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper considers the problem of selecting which ports to activate in a fluid antenna array (FAA) given a fixed number of RF chains. The authors propose IA-ZPA, which combines a compact CNN that scores ports based on CSI, a deterministic checkerboard zone projection that enforces per-zone port quotas and minimum spacing, and an inference-time mutual-impedance kernel to break ties. The selected mask is then evaluated by a current-domain regularized zero-forcing (RZF) precoder under accepted-power, current, and source-voltage constraints. In simulations with 256 candidate ports, 64 active ports, and 16 users over 500 channels, IA-ZPA achieves 76.28 bit/s/Hz constrained sum rate with mean PSLL -13.66 dB and 2.73 ms decision time, while the greedy baseline achieves 76.75 bit/s/Hz at -10.24 dB PSLL with 199.74 ms. The main claim is that IA-ZPA offers the best rate among methods satisfying the -13.5 dB mean-PSLL target, with much lower latency.","tokens_in":9545,"tokens_out":5990,"duration_ms":51882,"significance":"If the central claim holds, the paper contributes a practical, low-latency algorithm for FAA port activation that jointly accounts for rate, aperture geometry, and electrical feasibility, and it introduces a common current-domain evaluation protocol that makes selectors comparable under realistic source-drive limits. The module attribution in Table II is honest: it shows that the learned scoring and checkerboard projection, rather than the mutual-impedance kernel, drive most of the performance. However, the significance is conditional on demonstrating that the training-time proxy Rsel ranks masks consistently with the evaluation-time current-domain rate REM, which the paper does not yet establish.","major_comments":[{"comment":"The headline \"largest constrained rate\" claim depends on the ordering between the training-time proxy Rsel(a) and the evaluation-time constrained rate REM. The paper does not report any correlation (per-channel or pooled) between these two functionals, nor does it show that a different mask satisfying the -13.5 dB target cannot achieve a higher REM. This gap is made concrete by Table II, which shows that the inference-time mutual-impedance kernel changes REM by only 0.11 bit/s/Hz; almost all of IA-ZPA's gain therefore comes from the Rsel-trained CNN/checkerboard component. I request a correlation analysis and an additional PSLL-constrained greedy baseline (or a scorer retrained against REM) to test proxy alignment.","section":"Section III-B, Eq. (2) and Section IV-B, Table I"},{"comment":"The zero-shot port and user sweeps are reported using the normalized RZF-rate proxy Rsel, not the current-domain REM used in Table I. Because Rsel is also the training objective, these figures only show how the learned selector behaves under its training metric and do not provide evidence of load sensitivity for the paper's central constrained-rate result. The text in Section IV-E says the proxy is used \"to isolate the effect\", but this needs a justification that Rsel and REM respond similarly to load changes.","section":"Section IV-E, Figs. 3-4"},{"comment":"PSLL is computed from an equal-amplitude array factor, whereas the compared precoders are current-domain RZF with non-uniform amplitudes and a common scaling factor gamma. The paper acknowledges this in Section IV-F, but the central fitness target of -13.5 dB and all of Table I are expressed in equal-amplitude terms. It would strengthen the feasibility claim to show, for at least a few representative masks, that the equal-amplitude PSLL approximates the realized-radiation PSLL (or to argue why the equal-amplitude metric is the appropriate design constraint).","section":"Section II-A, Eq. (3) and Table I"},{"comment":"The 500-channel means are presented without error bars or significance tests. The differences between Greedy and IA-ZPA (0.47 bit/s/Hz) and between IA-ZPA and Uniform (1.36 bit/s/Hz) could be within channel variability; given that the paper's main claim is comparative, standard errors or confidence intervals are needed. Additionally, the absence of code or data limits verification of the numerical results.","section":"Table I"}],"minor_comments":[{"comment":"There is a stray comma in the argmax expression (\"{ zn - ... } ,\") that should be removed.","section":"Section III-C, Eq. (10)"},{"comment":"The table caption says \"Decision is the three-repeat median over 500 channels; one checkpoint is used per IA-ZPA call and results are pooled.\" The phrasing \"three-repeat median\" is unclear; it should be explained whether the median is over three checkpoint repetitions or over channels.","section":"Section IV-B, Table I note"},{"comment":"The y-axis lists methods but has no axis label; adding a label such as \"Method\" would improve clarity.","section":"Figure 1"},{"comment":"The caption of Fig. 2 says the realization is \"selected by a fixed above-median-rate and mean-PSLL-proximity rule,\" which is not described in the text; please clarify or omit this selection rule.","section":"Section IV-D, Fig. 2 caption"},{"comment":"The notation B^(c)_b for parity-c candidate sets is introduced, but the parity class is described only in prose; a short equation defining the parity would be helpful.","section":"Section III-C"},{"comment":"In the abstract, \"FIuid\" appears to be a typographical error for \"Fluid\".","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to require one more round of experiments to address the proxy-alignment issue. I would encourage the editor to ask for the code/data release or at least the correlation analysis between Rsel and REM."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Readable, honest engineering paper. The new combination is the checkerboard zone projection plus the inference-time mutual-impedance tie-break, evaluated under a shared current-domain induced-EMF backend. The ablation in Table II is the best part: it shows the impedance kernel the title advertises contributes only 0.11 bit/s/Hz, with CNN scoring and the checkerboard doing essentially all the work. That is the right way to report module attribution.\n\nThe simulation protocol is also sound. All selectors share the same 500 channels, drive budgets, noise calibration, and backend. Decision time excludes CSI acquisition and the common RZF stage, which is fair. Zero-shot sweeps are explicitly labeled as using the RZF proxy, separate from the current-domain evaluation, so there is no bait-and-switch in the figures.\n\nSoft spots: no code or data, and no error bars on the 500-channel means. The decisive comparisons are within one bit/s/Hz, and without variance bounds the ordering could easily flip on another seed. The bigger conceptual issue is the training proxy. Rsel in Eq. (2) ignores mutual coupling, amplitude weighting, and drive limits; the evaluation uses REM from Eqs. (11)-(14). The paper never tests whether these two functionals rank masks the same way. Table II makes this more urgent, not less: if the coupling kernel only changes REM by 0.11 bit/s/Hz, then the CNN/checkerboard, trained purely on Rsel, is responsible for the gains. Greedy also uses Rsel and gets 76.75 bit/s/Hz with worse PSLL, which is reassuring, but a per-channel rank correlation or a scatter plot between Rsel and REM is absent. Without it, the headline number is a property of this training run. Also, the delta over the authors' own learned blockwise activation [27] is not itemized, and the PSLL metric remains equal-amplitude rather than realized radiation (the paper acknowledges this as a next step).\n\nThis is a within-subfield contribution for FAA researchers. It deserves peer review, but only with a request for code/data or error bars and an explicit Rsel-to-REM alignment check. The authors are clearly capable of doing that; Table II shows they know what evidence looks like.","headline":"Honest ablation and clean protocol, but the load-bearing proxy alignment between training-time Rsel and evaluation-time REM is never tested.","tokens_in":10151,"tokens_out":2863,"would_cite":true,"duration_ms":27901,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A learned port picker for fluid antenna arrays beats greedy on rate, sidelobes, and speed.","keywords":["Fluid antenna array","Port activation","Mutual impedance","Current-domain precoding","Regularized zero forcing","Sidelobe suppression","Gumbel-Softmax","RF-chain budget"],"falsifier":"Retrain or rescore the same masks using the full induced-EMF constrained rate REM as the training signal instead of the proxy Rsel(a), and re-run the Table I comparison; if the rate margin over Uniform shrinks or Greedy becomes the best method that meets the -13.5 dB target, the proxy ordering is the reason for the gains. A second test is to tighten the port grid below the $\\lambda/3$ spacing enforced by the checkerboard and check whether the inference-time kernel penalty changes the selected mask enough to move REM by more than the 0.11 bit/s/Hz seen here.","tokens_in":8891,"feed_emoji":"📡","tokens_out":7837,"duration_ms":72420,"temperature":0.7,"pith_summary":"Fluid antenna arrays let a small number of radio-frequency chains choose which of many radiating ports to activate, and that choice simultaneously sets the multiuser channel and the sparse aperture. The paper claims that a compact learned scorer, combined with a deterministic checkerboard projection and an inference-time mutual-impedance penalty, can balance channel rate, aperture sidelobe level, RF-chain budget, and source-drive feasibility in one online rule. Under a common induced-EMF current-domain protocol over 500 channels, the proposed method satisfies the prescribed mean sidelobe target of -13.5 dB and reaches 76.28 bit/s/Hz, the largest constrained rate among methods meeting that target. The same rule makes the mask decision in 2.73 ms, compared with 199.74 ms for greedy selection, while greedy's rate advantage of 0.47 bit/s/Hz comes at a sidelobe level of -10.24 dB that misses the target. If the claims hold, real-time channel-adaptive port activation is feasible without sacrificing aperture quality or overloading the source drives.","feed_headline":"Learned antenna port picker beats greedy on rate, sidelobes, speed","feed_subtitle":"73x faster than greedy, meets the -13.5 dB sidelobe target, and gives up only 0.47 bit/s/Hz of rate.","key_machinery":"The load-bearing object is the IA-ZPA activation pipeline, whose three stages separate CSI preference from feasibility. A compact convolutional scorer maps eight standardized per-port features (coordinates, log field means and standard deviations, a coherence proxy, and log per-user powers) to a port score; a zone-wise Gumbel-Softmax relaxation with straight-through gradients trains the scorer against the normalized RZF-rate proxy plus a smooth sidelobe penalty, while the forward pass uses a hard top-$m_b$ mask so each of the $8 \\times 8$ zones contributes exactly one port. At inference, a raster-ordered choice over the two checkerboard parity lattices adds a mutual-impedance penalty $C_{mn}$ from the induced-EMF kernel of parallel half-wave dipoles and keeps the mask feasible; the better of the two parity masks is deployed. A shared current-domain RZF backend with accepted-power, current-norm, and source-voltage caps then computes the constrained rate $R_{\\mathrm{EM}}$ and the scaling factor $\\gamma$, so every comparison in Table I uses the same electrical model.","core_discovery":"The central claim is that rate, aperture quality, and electrical feasibility can be handled by a two-part activation rule: a learned CSI-conditioned port score provides the channel preference, while a zone-wise checkerboard projection fixes the active aperture and enforces exactly one port per zone, and a coupling-aware tie-break penalizes strongly mutually coupled choices at inference. All selectors are then evaluated by one current-domain regularized zero-forcing backend under accepted-power, total-current, and source-voltage limits. In the common 500-channel induced-EMF comparison, the full IA-ZPA rule attains REM = 76.28 bit/s/Hz with mean PSLL = -13.66 dB and median mask-decision time 2.73 ms, which is the largest constrained rate among the methods that meet the preselected -13.5 dB mean-PSLL target; greedy reaches 76.75 bit/s/Hz but at -10.24 dB PSLL and 199.74 ms. The attribution study isolates where the gain comes from: the learned scoring and the checkerboard projection contribute most of the rate-aperture tradeoff, while the mutual-impedance kernel changes the rate by only 0.11 bit/s/Hz and mostly lowers the average pairwise coupling measure from 0.02731 to 0.02718.","pith_inferences":["Editorial extension: the same learned-scores-plus-deterministic-projection split could be applied to other constrained combinatorial beamforming choices, such as reconfigurable-intelligent-surface element selection, where exact feasibility constraints, not score quality, are the bottleneck.","At the minimal $\\lambda/3$ spacing enforced by the checkerboard, mutual coupling appears to be a second-order effect on rate; on denser grids or with widerband excitation the kernel penalty's influence on the chosen mask could become first-order, a testable extension of the paper's setting.","Training the scorer on the full induced-EMF current-domain rate instead of the normalized RZF proxy would directly probe whether the proxy's rank ordering is why the gains transfer; Table II's small kernel effect suggests the proxy is adequate here, but the paper does not run that experiment.","The zero-shot sweeps use the normalized proxy, so a current-domain sweep over active-port count and user count would be needed to check whether the load behavior of Figs. 3 and 4 survives the induced-EMF backend."],"forward_implications":["Channel-adaptive port activation can meet a hard sidelobe aperture target without sacrificing rate: IA-ZPA improves constrained rate over the fixed Uniform mask by 1.36 bit/s/Hz while lowering mean PSLL by 3.40 dB.","The online decision time of 2.73 ms, against 199.74 ms for greedy, makes per-slot mask reconfiguration practical for real-time beamforming under the modeled 256-port, 64-active-port setup.","Because the zone-wise projection enforces the RF-chain budget exactly regardless of the learned scores, the feasibility guarantees do not depend on the scorer being perfect.","Under the same aggregate drive budget, dense full-port RZF hits the source-voltage limit before its accepted-power and current budgets, so a distributed 64-port sparse mask can outperform the dense 256-port array in constrained rate in this model.","Most of the rate-aperture gain is attributable to scoring plus checkerboard projection; the inference-time coupling kernel is a small electrical refinement in the tested configuration, not the main source of gain."],"supporting_citations":[{"why":"Supplies the regularized zero-forcing precoder form that defines both the training proxy and the current-domain backend.","marker":"[28]"},{"why":"Provides the Gumbel-Softmax relaxation used for the zone-wise soft top-$m_b$ scoring during training.","marker":"[29]"},{"why":"Provides the straight-through gradient estimator used in the soft-mask forward pass.","marker":"[30]"},{"why":"Establishes finite-aperture fluid antenna array design and the mask-configuration view of the reconfigurable aperture.","marker":"[17]"},{"why":"Introduces fluid antenna systems and the port-selection paradigm on which the activation problem is built.","marker":"[2]"},{"why":"Supplies the mutual-coupling analysis background for the induced-EMF kernel and the antenna interaction model.","marker":"[21]"},{"why":"Motivates the impedance-aware treatment of fluid antenna arrays with electromagnetic-aware modeling.","marker":"[22]"},{"why":"Relates to peak sidelobe suppression in planar fluid antenna arrays, connecting to the PSLL objective.","marker":"[25]"}],"fun_headline_variants":["IA-ZPA: 73x faster port selection, meets sidelobe target","73x faster antenna port selection that respects PSLL","Learned port scorer hits PSLL target 73x faster than greedy","Impedance-aware zonal activation: 73x speedup, meets sidelobe","Port activation that's 73x faster and still meets PSLL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the learned scorer, trained only on the normalized RZF-rate proxy Rsel(a) that ignores mutual coupling, sidelobe level, and source-drive limits, ranks masks in the same order as the full induced-EMF current-domain constrained rate; if that rank ordering diverges, the reported gains would not transfer.","fun_headline_variants_meta":{"raw":{"variants":["IA-ZPA: 73x faster port selection, meets sidelobe target","73x faster antenna port selection that respects PSLL","Learned port scorer hits PSLL target 73x faster than greedy","Impedance-aware zonal activation: 73x speedup, meets sidelobe","Port activation that's 73x faster and still meets PSLL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000802,"raw_usage":{"total_tokens":3544,"prompt_tokens":985,"completion_tokens":2559,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":2464}},"tokens_in":601,"tokens_out":2559,"duration_ms":19054,"temperature":1.0,"reasoning_tokens":2464,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:21:26.119420+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain or rescore the same masks using the full induced-EMF constrained rate REM as the training signal instead of the proxy Rsel(a), and re-run the Table I comparison; if the rate margin over Uniform shrinks or Greedy becomes the best method that meets the -13.5 dB target, the proxy ordering is the reason for the gains. A second test is to tighten the port grid below the $\\lambda/3$ spacing enforced by the checkerboard and check whether the inference-time kernel penalty changes the selected mask enough to move REM by more than the 0.11 bit/s/Hz seen here.","supporting_citations":[{"cited_title":"Large system analysis of linear precoding in corre- lated MISO broadcast channels under limited feedback,","cited_arxiv_id":null,"evidence_quote":"Supplies the regularized zero-forcing precoder form that defines both the training proxy and the current-domain backend."},{"cited_title":"Categorical reparameterization with Gumbel-Softmax,","cited_arxiv_id":null,"evidence_quote":"Provides the Gumbel-Softmax relaxation used for the zone-wise soft top-$m_b$ scoring during training."},{"cited_title":"Finite-aperture fluid antenna array design: Analysis and algorithm,","cited_arxiv_id":null,"evidence_quote":"Establishes finite-aperture fluid antenna array design and the mask-configuration view of the reconfigurable aperture."},{"cited_title":"Fluid antenna systems,","cited_arxiv_id":null,"evidence_quote":"Introduces fluid antenna systems and the port-selection paradigm on which the activation problem is built."},{"cited_title":"A review on array mutual coupling analysis,","cited_arxiv_id":null,"evidence_quote":"Supplies the mutual-coupling analysis background for the induced-EMF kernel and the antenna interaction model."},{"cited_title":"Electromagnetic-aware fluid antenna array,","cited_arxiv_id":null,"evidence_quote":"Motivates the impedance-aware treatment of fluid antenna arrays with electromagnetic-aware modeling."}],"review_version":1}