{"id":"6adcc668-ee3a-49a1-8886-5b7c13b6eb10","arxiv_id":"2507.14035","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A graph neural network with random port selection, accelerated on an FPGA, achieves competitive weighted sum-rate in fluid antenna MIMO systems.","lead":"This paper proposes a way to choose antenna ports and compute beamforming in fluid antenna systems using a graph neural network, then designs a custom FPGA chip to make the computation fast. If it works, it could make reconfigurable fluid antennas practical in wireless networks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The GNN input features omit all inter-cell channel information (Eq. 6), so the claimed multi-cell WSR gains cannot stem from the inter-cell interference management the paper invokes.","rationale":"I examined the central claim as stated: the GNN-RPS approach jointly optimizes port selection and beamforming in a multi-cell FAS network, achieving WSR gains over baselines. For that claim to hold, each BS's beamforming must account for interference it causes to other-cell UEs, as the WSR (P1) includes cross terms h_{ik,j}w_{jr}. Eq. (6) defines the only input to the i-th GNN as the desired user channels h_{ik,i}; no cross-cell channels h_{jk,i} are provided. Thus the GNN output is a function of local CSI only, making the asserted 'inter-cell interference pattern learning' (Sec. IV-A4) and 'network-wide optimization through distributed cooperation' unattainable. Training on a global loss cannot compensate for absent input information; it can only fit an average policy. This is a correctness risk internal to the paper, not a disagreement with field consensus. The reader's quantization concern is legitimate but secondary: even with perfect fixed-point accuracy, the communication results are compromised by this architectural mismatch. I therefore view the conditional verdict as appropriate, with the additional condition that the authors either include cross-cell CSI in the GNN input or substantially reframe the contribution as local-only beamforming. The proposed concrete test—retraining with augmented cross-cell features—would settle the matter.","tokens_in":16327,"tokens_out":7161,"duration_ms":92557,"concrete_test":"Retrain the model with an augmented input that concatenates, for each cell i, the cross-cell channels {h_{jk,i} : j != i, all k} as additional node or edge features, keeping all other training details identical. If the augmented GNN achieves materially higher WSR (e.g., >5%) than the paper's local-CSI GNN on the same test set, the current architecture's performance claims rest on omitted information and the inter-cell management claim fails. As a complementary analytical check, for I=2 and K=1, derive the optimal w_i and show it depends on h_{21,1}; because X_in_1 contains only h_{11,1}, the GNN cannot represent that optimum.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is the mismatch between the problem formulation and the GNN input. The WSR objective (P1) depends on inter-cell interference: the SINR in (5) includes terms |h_{ik,j} w_jr|^2 from other BSs. Therefore, to choose w_i for cell i that maximizes global WSR, the beamformer must depend on cross channels h_{jk,i} (channels from BS i to UEs in other cells). However, the node feature matrix X_in_i in Eq. (6) contains only Re/Im of h_{ik,i}, the desired channels from BS i to its own UEs. No inter-cell channel information is supplied to any GNN. The model thus computes w_i as a function solely of local desired CSI. Consequently, the claim in Sec. IV-A4 that the multi-GNN 'effectively learns to characterize both inter- and intra-cell interference patterns' and enables 'network-wide performance optimization through distributed cooperation' is not realizable: the output is invariant to the instantaneous realization of the interference links it is supposed to suppress. Training on a global WSR loss cannot conjure information that is absent from the input; at best it learns a policy optimized in expectation over the channel distribution. The WSR gains in Figs. 6-10 therefore cannot be attributed to inter-cell interference management, and the comparison to MMSE/ZF/MRT baselines is not a comparison of coordinated multi-cell beamforming. This is an internal inconsistency between the stated objective and the architecture.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hardware-software co-design framework for fluid antenna systems (FASs). It formulates a weighted-sum-rate (WSR) maximization problem for an FA-enabled downlink multi-cell MIMO network with port selection and beamforming, proposes a multi-GNN architecture with random port selection (GNN-RPS) to solve it, and implements the GNN inference on an FPGA with a custom instruction set, systolic-array micro-architecture, and a scheduling algorithm for concurrent inference. Simulations compare GNN-RPS against MMSE, ZF, and MRT baselines under random port selection, and FPGA experiments report inference latency for single and multiple concurrent port-selection tasks.","tokens_in":16665,"tokens_out":7156,"duration_ms":93670,"significance":"If fully supported, the paper would be a useful step toward practical FAS by showing that a learned beamformer combined with random port selection can approach the performance of much larger search budgets, and by demonstrating that multiple GNN inference tasks can share an FPGA accelerator with nearly unchanged latency. The scheduling result in Section VI-C, where four concurrent inferences run almost as fast as one, is an interesting and clearly reported hardware contribution. The central communication claim, however, is weakened by the mismatch between the inter-cell objective and the local-only GNN input features, and by the absence of an optimization-based baseline such as WMMSE. The hardware claim also lacks quantization-fidelity evidence for the 8-bit fixed-point implementation. With those points addressed, the paper could make a solid contribution to system-level FAS design.","major_comments":[{"comment":"The node feature matrix X_in_i in Eq. (6) contains only Re/Im of the desired channels h_{ik,i} for UEs in cell i; no cross-cell channels h_{ik,j}, j≠i, are provided to any GNN. Because the SINR in Eq. (5) and hence the WSR objective (P1) depend on inter-cell interference terms |h_{ik,j}w_{jr}|^2, a beamformer computed from X_in_i is invariant to the instantaneous realizations of those interference links. The claim in Sec. IV-A4 that the multi-GNN 'effectively learns to characterize both inter- and intra-cell interference patterns' and enables 'network-wide performance optimization through distributed cooperation' is therefore not supported by the architecture. Training on a global WSR loss cannot introduce information that is absent from the input; at best the model learns a policy optimized in expectation over the channel distribution. Please either incorporate cross-cell channels into the input features or revise the claims and interpret Figs. 7-10 as gains of a per-cell policy rather than coordinated inter-cell beamforming.","section":"Sec. IV-A, Eq. (6), Sec. IV-A4"},{"comment":"The conclusion that GNN-RPS achieves 'competitive communication performance' is drawn only against MMSE, ZF, and MRT baselines. Since P1 is a standard WSR maximization problem, a WMMSE-type optimization baseline using the same random port selections and the same CSI is the natural reference. Without such a baseline, the reader cannot judge how close the learned beamformers are to the optimum, or whether the reported gains reflect the GNN or merely the relative weakness of the baselines. Please add a WMMSE or other state-of-the-art optimization-based baseline.","section":"Sec. VI-B, Figs. 7-10"},{"comment":"The FPGA experiments use 8-bit fixed-point arithmetic and state that this gives 'only negligible loss in accuracy,' but no comparison is provided between 8-bit and floating-point beamforming outputs or WSR values. Since all latency numbers in Sec. VI-C are obtained with this representation, the hardware-software co-design claim requires evidence that quantization does not materially change the communication performance reported in Sec. VI-B. Please report quantization-aware simulation results or a hardware-in-the-loop comparison.","section":"Sec. VI-C"},{"comment":"The output normalization Xout_i ← √P LayerNorm(X_i^(8)) is not shown to enforce the per-BS power constraint C1. Standard LayerNorm normalizes each row (or the full matrix) to zero mean and unit variance, so the resulting total transmit power is not necessarily equal to P and may violate C1 when K_i>1. Please specify the exact normalization operation, state how the power constraint is enforced, and confirm that the WSR values in Figs. 6-10 are computed with feasible beamformers.","section":"Sec. IV-A5, Algorithm 1 line 19"}],"minor_comments":[{"comment":"The scheme labeled 'GNN-Exhaustive' uses 500 independent random trials, which is not an exhaustive search over all L^N port combinations; the label and the associated 'upper bound' claim should be revised to avoid overstating the reference.","section":"Sec. VI-B"},{"comment":"The complexity notation L1,...,L8 is not mapped to the dimensions given in Table II; please define these quantities or align them with the table so that the complexity expressions can be verified.","section":"Sec. IV-A5"},{"comment":"Equation (1) appears to have a typesetting error in the Bessel-function argument, with a missing closing parenthesis; please correct the formula.","section":"Eq. (1)"},{"comment":"In Algorithm 1, lines 5-9 compute MLP1(x_i,k') repeatedly for every k, although the MLP output does not depend on k; the text should clarify whether this is intentional or a simplified description, since it affects the stated O(K_i^2) complexity.","section":"Algorithm 1"},{"comment":"The manuscript should state explicitly how the channel vector for a selected port is obtained from the correlated model in Eqs. (1)-(3), since this is needed to reproduce the RPS simulations.","section":"Sec. III"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the hardware demonstration is a genuine strength. The main blockers are the inconsistency between the inter-cell objective and the local-only GNN inputs, the missing WMMSE baseline, and the unverified quantization claim. These are fixable in revision, so I do not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a legitimate, well-executed incremental contribution: it applies GNN-based beamforming to fluid antenna port selection and builds an FPGA accelerator with a scheduling algorithm that lets four port-selection inferences run almost at single-inference latency. The RPS analysis (70% of the 2000-trial WSR with only 20 trials) is useful engineering knowledge, and the hardware measurements are real experiments, honestly caveated as still exceeding typical coherence times. The authors are clearly fluent in both the FAS literature and the GNN-beamforming literature; the writing is straightforward and the approach is reproducible in simulation from the described details.\n\nThe biggest soft spot is the mismatch the stress-test note flags: the GNN inputs (Eq. 6) contain only the desired channels from each BS to its own UEs, no inter-cell channels. The multi-cell WSR objective depends on cross links, so the trained GNN cannot adapt to instantaneous inter-cell interference. Claims in Sec. IV-A4 that the model \"effectively learns to characterize both inter- and intra-cell interference patterns\" and enables \"network-wide performance optimization through distributed cooperation\" are not supported by the architecture. At best, centralized training over a global WSR loss learns a policy that is robust to interference in expectation, but the inference is per-cell and local. This is an overstatement rather than a fatal flaw: the baselines (MMSE, ZF, MRT) also use local CSI, so the comparisons are fair, but the paper should not claim coordinated multi-cell beamforming gains.\n\nSecond, the absence of a WMMSE or other optimization-based baseline weakens the \"competitive performance\" claim. GNN beating ZF/MMSE/MRT is not a high bar; a WMMSE comparison would calibrate how much is lost by the learned local policy. Third, the 8-bit fixed-point FPGA results are presented with latency only, no resource/power, and no quantified accuracy loss in beamforming output or WSR, so the link between hardware speed and communication performance is unverified. Fourth, no code or data are released; given the heavy self-citation and the incremental nature, community access to code would materially strengthen the paper.\n\nThe citation pattern is heavy on the authors' own FAS work, but that is common in a young subfield where they are central; it is not a disqualifier. The central contribution is new, clearly described, and supported by real experiments. This deserves a serious referee, but the authors should be asked to either add inter-cell CSI to the GNN input and rerun the simulations, or substantially tone down the inter-cell interference claims. The hardware section is fine and could be a separate contribution on its own. For a reader in FAS or GNN-based beamforming, this is worth reading and citing once the claims are aligned with the architecture.","headline":"Solid but incremental FAS engineering: the GNN-RPS + FPGA co-design is plausible and the hardware scheduling is genuinely useful, but the paper overclaims inter-cell interference management when its GNN inputs are strictly local CSI.","tokens_in":17146,"tokens_out":1770,"would_cite":false,"duration_ms":25097,"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 graph neural network with random port selection, accelerated on an FPGA, makes fluid-antenna beamforming and port selection practical.","keywords":["fluid antenna systems","port selection","beamforming","graph neural networks","random port selection","FPGA acceleration","weighted sum rate maximization","hardware-software co-design"],"falsifier":"Run the trained GNN on the FPGA in 8-bit fixed point and in a higher-precision reference using the same channel samples as the simulations, and compare the output beamforming vectors or the resulting weighted sum rate; a non-negligible gap would break the link between the measured latency and the claimed communication performance.","tokens_in":16182,"feed_emoji":"📡","tokens_out":9571,"duration_ms":95065,"temperature":0.7,"pith_summary":"This paper tries to establish that fluid antenna systems can be made practical by co-designing the optimization algorithm with the hardware that runs it. The algorithmic claim is that a multi-cell graph neural network (GNN) for beamforming, paired with random port selection, reaches near-exhaustive weighted-sum-rate performance after only tens of random port choices and outperforms minimum mean square error (MMSE), zero-forcing (ZF), and maximum ratio transmission (MRT) baselines. The hardware claim is that an instruction-driven field-programmable gate array (FPGA) accelerator can run several of these port-selection beamforming inferences concurrently with little added latency. If both claims hold, the computationally prohibitive mixed-integer optimization of port selection and beamforming can be replaced by a deployable real-time pipeline.","feed_headline":"Random port picks plus GNN match near-exhaustive fluid-antenna gains","feed_subtitle":"A graph neural network and FPGA co-design run up to four port selections with almost no extra latency.","key_machinery":"The load-bearing mechanism is a multi-GNN architecture with centralized training and distributed inference, in which each base station runs a homogeneous GNN module that maps local channel state information, the real and imaginary parts of its channels to the served users, into beamforming vectors. Each GNN consists of an input multi-layer perceptron (MLP), two GNN layers whose aggregation is an MLP followed by max-pooling over the other users and whose combination is concatenation followed by an MLP, and a final fully connected layer with power-normalizing LayerNorm. Port selection is handled by random port sampling, keeping the best of several trials. On the FPGA, the key mechanism is an instruction-driven overlay accelerator with a custom instruction set architecture, systolic-array matrix units, ping-pong double buffering, and a scheduling algorithm that concatenates inputs from multiple port selections, reuses shared weights and intermediate results, and reorders computation so that memory-bound GNN tasks overlap instead of stalling.","core_discovery":"The paper claims that the mixed-integer problem of jointly choosing fluid-antenna ports and computing downlink beamforming vectors can be split into a GNN for beamforming and random port selection for port choice. In the simulated two-cell, four-user-per-cell setting, 20 random port selections reach about 70% of the weighted sum rate obtained with 2,000 selections, and the GNN-based scheme beats MMSE, ZF, and MRT at every transmit power tested. On the hardware side, an instruction-driven FPGA accelerator with a custom instruction set and a scheduling algorithm that shares model parameters and reuses intermediate results executes up to four concurrent port-selection inferences in 3.994 to 6.222 ms, nearly the same as the 3.926 to 6.104 ms needed for a single inference. The paper presents the combination as a complete hardware-software path toward real-time intelligent beamforming in fluid antenna systems.","pith_inferences":["If the 8-bit fixed-point quantization truly costs negligible accuracy, the same scheduling strategy should scale beyond four concurrent port selections once the workload shifts from memory-bound to compute-bound, so adding on-chip compute or sparsity could extend the parallelism further.","The RPS result suggests port selection can be treated as a cheap stochastic search; a learned selection policy trained alongside the GNN could capture additional gains over random trials at modest extra inference cost.","Because the accelerator is instruction-driven and overlay-based, the same hardware design could be retargeted to other GNN-based wireless optimization problems, such as RIS-assisted or cell-free beamforming, by recompiling instructions rather than redesigning the chip.","A direct precision comparison on the FPGA would be the natural next experiment; until it is reported, the communication-performance claims and the latency claims rest on separate evidence."],"forward_implications":["Random port selection with tens of trials can approximate the performance of a far larger port search, so the combinatorial port-selection bottleneck in fluid antenna systems is much smaller than exhaustive enumeration suggests.","GNN-based beamforming can outperform MMSE, ZF, and MRT in multi-cell fluid-antenna downlinks because the multi-GNN learns inter-cell and inter-user interference patterns from channel features.","Because the accelerator reuses loaded model weights and intermediate results, the latency of processing several port selections is close to that of one, up to about four concurrent tasks.","Measured inference latencies of about 3.9 to 6.2 ms provide a prototype baseline for real-time beamforming, with model compression and ASIC implementations expected to push latency lower."],"supporting_citations":[{"why":"Introduces the fluid antenna system whose port-switching degrees of freedom the optimization problem exploits.","marker":"[7]"},{"why":"Supplies the spatial-correlation channel model used to generate correlated ports in the simulations.","marker":"[32]"},{"why":"Provides the multi-GNN architecture for weighted-sum-rate maximization that the proposed design adapts to port selection.","marker":"[19]"},{"why":"Establishes that a GNN can directly output beamforming vectors for sum-rate maximization in multiuser systems.","marker":"[21]"},{"why":"Provides the FPGA overlay processor methodology that the instruction-driven accelerator extends to the GNN workload.","marker":"[31]"},{"why":"Supplies the roofline analysis showing the workload is memory-bound, which motivates the task scheduling algorithm.","marker":"[35]"}],"fun_headline_variants":["GNN plus random ports: near-exhaustive gains at 70% effort","Fluid-antenna co-design: GNN and FPGA cut beamforming latency","Random port picks and GNN rival exhaustive search in FAS","FPGA + GNN make fluid-antenna beamforming real-time","Four port selections in one inference: FAS hardware-software win"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the FPGA's 8-bit fixed-point arithmetic loses negligible beamforming accuracy; if quantization materially degrades the beamforming vectors, the measured low latency would not reproduce the communication performance shown in simulation.","fun_headline_variants_meta":{"raw":{"variants":["GNN plus random ports: near-exhaustive gains at 70% effort","Fluid-antenna co-design: GNN and FPGA cut beamforming latency","Random port picks and GNN rival exhaustive search in FAS","FPGA + GNN make fluid-antenna beamforming real-time","Four port selections in one inference: FAS hardware-software win"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1270,"prompt_tokens":920,"completion_tokens":350,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":255}},"tokens_in":536,"tokens_out":350,"duration_ms":4454,"temperature":1.0,"reasoning_tokens":255,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:11:03.100876+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained GNN on the FPGA in 8-bit fixed point and in a higher-precision reference using the same channel samples as the simulations, and compare the output beamforming vectors or the resulting weighted sum rate; a non-negligible gap would break the link between the measured latency and the claimed communication performance.","supporting_citations":[{"cited_title":"Flui d antenna systems,","cited_arxiv_id":null,"evidence_quote":"Introduces the fluid antenna system whose port-switching degrees of freedom the optimization problem exploits."},{"cited_title":"An information-theoretic characterization of MIMO-FAS: Optimization, diversity-multiplexing tradeoff and q-outage capacity,","cited_arxiv_id":null,"evidence_quote":"Supplies the spatial-correlation channel model used to generate correlated ports in the simulations."},{"cited_title":"A distributed machine learning-based approach for IRS- enhanced cell-free MIMO networks,","cited_arxiv_id":null,"evidence_quote":"Provides the multi-GNN architecture for weighted-sum-rate maximization that the proposed design adapts to port selection."},{"cited_title":"GNN-based beamforming for sum-rate maximization in MU-MISO networks,","cited_arxiv_id":null,"evidence_quote":"Establishes that a GNN can directly output beamforming vectors for sum-rate maximization in multiuser systems."},{"cited_title":"Graph-OPU: A highly flexible FPGA-based overlay pro- cessor for graph neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the FPGA overlay processor methodology that the instruction-driven accelerator extends to the GNN workload."},{"cited_title":"Optimizing FPGA-based accelerator design for deep convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the roofline analysis showing the workload is memory-bound, which motivates the task scheduling algorithm."}],"review_version":1}