{"id":"3a22a0bc-78bb-44b3-ab25-b123ea09d6f3","arxiv_id":"2411.15762","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"GGML uses gradient-guided meta-learning with KAN networks to alternately optimize analog and digital precoders in mmWave MU-MISO systems, outperforming conventional hybrid precoding baselines in simulations.","lead":"A new algorithm called GGML applies meta-learning and Kolmogorov-Arnold networks to hybrid precoding in millimeter-wave multi-user systems, claiming higher spectral efficiency and up to 8 times faster runtimes than existing methods. The paper may interest wireless engineers looking for low-complexity, robust precoding without offline training.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'outperforms fully digital WMMSE' claim (Fig. 6c) is the load-bearing risk: since any hybrid F·D is a feasible fully digital precoder, a correctly run WMMSE cannot be beaten; the baseline's RF-chain count, stopping rule, and normalization are unspecified, so the headline comparison may be an…","rationale":"The reader's conditional verdict is appropriate, but the weakest assumption they identified (online-KAN convergence with random initialization) is not the most load-bearing concern. Convergence can be checked empirically and is already supported by the provided convergence plots. The stronger, theory-adjacent risk is the WMMSE comparison: because every hybrid precoder F D is a feasible fully digital precoder, the optimal fully digital precoder dominates the optimal hybrid precoder. Thus the claim that GGML outperforms fully digital WMMSE is only plausible if the WMMSE baseline was not the fully digital optimizer or was mis-specified. The reader's rationale does note that baseline configurations, notably WMMSE, are unspecified, so there is partial agreement; however, elevating this from a specification gap to the central objection changes which experiment should be run first. The paper's code link helps: running the WMMSE script and inspecting the baseline's RF-chain count and stopping rule would settle the concern quickly. This does not require rejecting the paper, but it does require the authors to clarify and re-verify the WMMSE baseline before the headline claims can be accepted. The verdict therefore stays conditional.","tokens_in":33,"tokens_out":6352,"duration_ms":212025,"concrete_test":"Reproduce Fig. 6c with the standard fully digital WMMSE algorithm of [15] under exactly the same system settings: N antennas, N RF chains, K single-antenna users, total power P=1, same channel realizations and SNR, and the same user weights. Run WMMSE to a tight convergence tolerance (e.g., relative objective change < 1e-6) from several random and ZF initializations, keeping the best result. Then evaluate the objective for the GGML output W = F D, with D power-normalized as in (16). If the fully digital WMMSE value is ≥ the GGML value at every M, the headline claim fails; if GGML exceeds it, check whether the baseline used fewer than N RF chains or a different stopping/power rule, and report those settings explicitly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decisive issue is not KAN convergence but the WMMSE comparison in Section V-D. For any hybrid solution (F,D), the effective precoder is W = F D ∈ C^{N×K} with ||W||_F^2 ≤ P. A fully digital system with N antennas and N RF chains can realize exactly the same W, so the feasible set of the fully digital problem contains every hybrid precoder. Therefore the global maximum of the weighted sum-rate (or any fixed objective) over fully digital precoders is at least the maximum over hybrid precoders. Consequently, a correctly implemented and converged fully digital WMMSE (refs. [15], [42]) cannot be outperformed by GGML when M>6, or at any M. The sentence 'GGML even outperforms WMMSE precoding (simulated as a fully digital system with the same number of antennas)' is only possible if the WMMSE baseline is not actually the fully digital optimizer: e.g., it was run with M RF chains instead of N, stopped early, initialized poorly, or used a different power normalization. The paper gives no WMMSE initialization, convergence criterion, number of RF chains, or transmit-power handling for that baseline. Because this is the paper's most striking claim, the entire central comparison rests on this unspecified baseline. The valid conclusion would be either that WMMSE was misconfigured (making the comparison unfair), or that the paper has found a violation of a simple feasibility upper bound, which is implausible without further evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GGML, a gradient-guided meta-learning framework for hybrid analog-digital precoding in millimeter-wave MU-MISO systems. The digital and analog precoders are updated alternately using neural networks (KAN-based) that take the gradient of the objective as input, while the network parameters are updated globally with a loss that includes a variance penalty. The framework is extended to imperfect CSI by transforming a worst-case maximin problem into a geometric-program-like MMSE minimization. Simulation results claim that GGML outperforms manifold optimization, majorization-minimization, and CNN baselines, converges faster, and even surpasses fully digital WMMSE precoding when the number of RF chains exceeds 6.","tokens_in":17647,"tokens_out":6353,"duration_ms":57931,"significance":"If the claims were substantiated, the approach would offer a low-complexity, training-free hybrid precoding method that also handles imperfect CSI, which is of practical interest for mmWave systems. The paper provides a worst-case reformulation with a proof, a complexity analysis, and a simulation code link, which are strengths. However, the headline result that a hybrid scheme outperforms fully digital WMMSE is mathematically questionable, the loss function is not fully specified, and the statistical evidence lacks error bars and sensitivity analysis. The contribution is potentially useful but requires substantial revision and re-validation before the claims can be accepted.","major_comments":[{"comment":"The claim that GGML outperforms fully digital WMMSE with the same number of antennas is inconsistent with a basic feasibility argument. For any hybrid precoder (F,D) satisfying the power constraint, the effective linear precoder W = F D in C^{N x K} is also feasible for a fully digital system with N antennas and N RF chains. Therefore the optimal value of the fully digital weighted sum-rate problem is an upper bound on the hybrid optimum. A correctly implemented and converged WMMSE baseline (refs. [15] and [42]) cannot be outperformed. The paper does not report the RF-chain count, initialization, convergence criterion, or power normalization used for the WMMSE baseline. This headline comparison must be corrected or removed; as it stands, the most striking result of the paper is unsupported.","section":"Section V-D, Fig. 6c"},{"comment":"The loss function L = -R + beta * Var(R) is not well-defined because R in (5) is a scalar (the weighted sum rate), so Var(R) has no meaning unless it is defined as the variance of per-user rates R_k or over some other ensemble; no such definition is given. Moreover, the penalty coefficient beta is never specified in Section V, and no sensitivity analysis with respect to beta is reported. Since this loss directly drives the global update of the network parameters in Algorithm 1 (line 8) and Algorithm 2 (line 9), this ambiguity is load-bearing for reproducibility and for understanding the algorithm's behavior.","section":"Section III-B-3 and IV-B, Eqs. (19) and (34)"},{"comment":"The claim that GGML is 'free-of-training' and 'plug-and-play' (abstract and Section I-B) is contradicted by the algorithm itself. Lines 8-10 of Algorithm 1 and lines 9-11 of Algorithm 2 update the network parameters theta_D and theta_F using Adam on the same channel instance that is later used for evaluation. There is no pre-training phase, no task distribution, and no separate adaptation set. At best, this is per-instance online adjustment of a randomly initialized network; it is not meta-learning in the standard sense of learning across a distribution of tasks, and it requires online computation during deployment. The terminology should be corrected and the computational cost of these online updates should be included in the complexity claims.","section":"Algorithms 1 and 2"},{"comment":"The numerical results are averaged over 102 channel realizations, but no error bars, confidence intervals, or sensitivity analyses are reported. Some reported gains are small (e.g., 1.4% over MO at convergence in Section V-B), so without a measure of dispersion it is unclear whether the improvements are statistically meaningful. In addition, the KAN architecture (number of layers, width, grid size, spline order) and the Monte Carlo outage bound epsilon_k used for Fig. 8 are not disclosed. The paper provides a code link, which is good, but the manuscript itself should specify these essential simulation details to allow independent reproduction.","section":"Section V-A and Figs. 4-8"},{"comment":"The worst-case robust formulation (P2) assumes a bounded error set with sigma_max(e_k) <= epsilon_k, but the simulation model (Section V-A) draws e_k from an unbounded Gaussian distribution. The Monte Carlo outage construction only bounds the error with 5% probability, so the guarantee is probabilistic rather than the worst-case guarantee implied by the formulation. The paper should clarify whether the robust design is meant to satisfy the bound only with high probability, and it should report the outage level and the resulting epsilon_k values used in the figures.","section":"Section IV-A and V-A"}],"minor_comments":[{"comment":"The notation DPN is used both for the network and for its output; the text should clarify that DPN(.) denotes the neural network function, not the resulting precoding matrix.","section":"Section III-B-2, Eq. (15)"},{"comment":"The symbol e_k is used for both the channel estimation error in Eq. (8) and the MMSE error in Eq. (22); this is confusing and should be resolved by renaming one of them.","section":"Section II-B-2 and IV-A"},{"comment":"The parameter beta is described as a 'penalty rate'; it should be called a penalty coefficient, and its value in the simulations should be stated.","section":"Section III-B-3, Eq. (19)"},{"comment":"The caption should state the number of antennas N and the exact configuration of the WMMSE baseline (number of RF chains, iterations, initialization) so that the comparison is interpretable.","section":"Section V-D, Fig. 6c"},{"comment":"The sentence 'only requires deploying lightweight neural networks at the user end' appears to be a typo; the proposed algorithm runs at the base station, not at the user equipment, and should be corrected.","section":"Section VI, Conclusion"},{"comment":"The text mentions both '102 random initializations' and '102 independent channel realizations'; please clarify which quantity is averaged over in each figure.","section":"Section V-B"},{"comment":"The GGML complexity O(L N M K^2) does not include the complexity of the KAN forward and backward passes, which depends on the network width and number of layers; this should be acknowledged.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The WMMSE outperformance claim in Fig. 6c appears theoretically impossible for a correctly configured fully digital baseline, which suggests a possible baseline implementation error. The authors should be asked to provide the exact setup of the WMMSE comparison or to remove that claim. The paper also needs to clarify the 'free-of-training' terminology and disclose all hyperparameters. The topic fits the journal, but the current validation is not sufficient for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe core idea is fine: using a learned optimizer (a KAN) to guide alternating updates in hybrid precoding is a legitimate new application of the meta-learning framework from [35]–[37]. But the paper's headline claim—that GGML beats fully digital WMMSE when M > 6—is almost certainly an artifact of an under-specified baseline. And the 'free-of-training' framing is a stretch: the KAN is trained online on each channel instance, so it's per-instance adaptation, not meta-learning across tasks.\n\nWhat the paper does well: it adapts a known gradient-guided meta-learning approach to hybrid precoding, uses gradients rather than raw channels as network input, and extends the method to imperfect CSI via a worst-case MMSE transformation. The complexity analysis is provided, and the authors include a code link. That's real work.\n\nNow the soft spots, in proportion to how soft they are. The WMMSE comparison is the biggest issue. Any hybrid F·D is feasible as a fully digital precoder, so a well-tuned WMMSE should not be worse than a hybrid method at the same transmit power. The paper doesn't state WMMSE's initialization, number of iterations, or power normalization, so the comparison is impossible to verify. This needs to be fixed or the claim dropped. The 'free-of-training' claim is misleading: the networks are randomly initialized and then updated via Adam during the L iterations. That is not 'training-free' in any useful sense; it's just online adaptation. And the paper never gives β (the variance penalty coefficient) or any KAN architecture details (depth, width, grid size). No error bars or sensitivity analysis either, so the simulation results are hard to assess.\n\nThe derivations in Sections III and IV look correct to me, and the literature review is fair. If this is a student's first paper, it's a solid start, but it needs major revision before it's publishable.\n\nWho is this for? People working on ML-aided hybrid beamforming, especially those interested in learned optimizers. It's a reasonable conference submission after the claims are tightened and the baseline is specified.\n\nRecommendation: yes, it deserves serious peer review. A good referee will catch the WMMSE problem and the overclaim. Send it out, but expect heavy revision.","headline":"A plausible learning-aided hybrid precoding method, but the headline WMMSE comparison and 'free-of-training' framing are both overclaimed.","tokens_in":18156,"tokens_out":4484,"would_cite":false,"duration_ms":38581,"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 meta-learning loop can optimize hybrid precoders without pre-training, beating fully digital WMMSE.","keywords":["hybrid precoding","millimeter wave","MU-MISO","meta-learning","Kolmogorov-Arnold networks","spectral efficiency","imperfect CSI","alternating optimization"],"falsifier":"Run GGML on a larger or ill-conditioned system than those tested, such as N = 512 or K = 16 with near-degenerate scattering clusters, and check whether a fresh random seed still reaches the claimed spectral efficiency within 500 iterations; if some seeds stall below the manifold-optimization baseline or diverge, the free-of-training claim fails.","tokens_in":1844,"feed_emoji":"📡","tokens_out":2586,"duration_ms":68106,"temperature":0.7,"pith_summary":"This paper tries to establish that hybrid analog-digital precoding in millimeter-wave multi-user systems can be solved by a meta-learning loop that needs no pre-training: feed the raw gradient of the spectral-efficiency objective into a small Kolmogorov-Arnold network, use its output as the next update of the digital or analog precoder, and update the network parameters from a global loss. The claim is that this alternating scheme out-achieves classical optimization baselines and data-driven CNN precoders, converges in about 200 iterations, and can surpass fully digital WMMSE once the number of RF chains exceeds six. A robust extension handles imperfect channel knowledge by converting the worst-case max-min problem into a geometric program over an effective channel. If correct, the framework offers a plug-and-play, low-complexity alternative to training-based deep learning for hybrid beamforming.","feed_headline":"Meta-learning hybrid precoder beats fully digital WMMSE","feed_subtitle":"No pretraining: KAN-guided alternating updates beat traditional precoders and speed up by 8x.","key_machinery":"The central mechanism is the gradient-guided meta-learning loop with two KAN subnetworks, the digital precoding network and the analog precoding network. In each outer iteration, the gradient of the loss with respect to the current digital precoder is fed to one KAN to produce an update, $\\mathbf{D} = \\mathbf{D}_0 + \\operatorname{DPN}\\big(\\nabla_{\\mathbf{D}_0}\\mathcal{L}\\big)$, followed by power rescaling; likewise, the analog precoder is updated as $\\mathbf{F} = \\mathcal{P}_{\\mathcal{F}}\\big(\\mathbf{F}_0 + \\operatorname{APN}\\big(\\nabla_{\\mathbf{F}_0}\\mathcal{L}\\big)\\big)$, where $\\mathcal{P}_{\\mathcal{F}}$ projects onto unit-modulus entries. The KANs use learnable B-spline activation functions, which the paper argues capture subtle structure in the gradient flow. A global perspective is imposed by updating the KAN parameters with an adaptive-moment optimizer on the loss $\\mathcal{L} = -R + \\beta\\,\\mathrm{Var}(R)$, coupling the two subproblems and guiding the iterations away from local optima. For imperfect CSI, Lemma 1 bounds the uncertain channel Gram matrix and Proposition 2 shows that the worst-case MSE is attained at that bound, so the robust problem reduces to the same GGML loop with an effective channel.","core_discovery":"The paper claims that the non-convex hybrid precoding problem can be solved by gradient-guided meta-learning (GGML): instead of learning a mapping from channels to precoding matrices, two KAN-based networks consume the objective gradient with respect to the current precoder and output update directions, while the network parameters themselves are updated online from a global loss. Digital and analog precoders are optimized alternately, with the analog update projected onto the constant-modulus phase-shifter constraint and the digital update rescaled to full transmit power. The authors report that this approach attains higher average spectral efficiency than manifold optimization, majorization-minimization, and CNN baselines, converges within about 200 iterations, runs about eight times faster than manifold optimization at 256 antennas, and can exceed the spectral efficiency of fully digital WMMSE when the number of RF chains is larger than six. The same framework is extended to imperfect CSI by replacing the estimated channel with a worst-case effective channel, turning the maximin problem into a geometric program over worst-case mean-square error.","pith_inferences":["Beyond the paper's setting, the same gradient-fed alternating loop should apply to other non-convex constrained optimization problems whose variables admit projection operators, such as constant-modulus, quantized, or unit-norm beamforming.","Because the KAN parameters are trained online per channel, the method's robustness rests on the optimizer rather than on transferable features; a natural extension is warm-starting the KAN parameters from a previous channel to accelerate adaptation in time-varying channels.","The reported ability to beat fully digital WMMSE suggests that the learned update path acts as an implicit regularizer adapted to the phase-shifter constraint; verifying whether the gap persists on measured or ray-traced channels would strengthen the deployment claim.","A testable sensitivity question is how the fixed budget of 500 iterations and random network initialization interact: measuring seed-to-seed variance in spectral efficiency would show how much of the gain is structural versus initialization luck."],"forward_implications":["Hybrid precoders can be obtained with no offline training, since the KAN networks are freshly initialized and updated online for the given channel instance.","The per-iteration complexity scales as $O(LNMK^2)$, giving roughly eight times less CPU time than manifold optimization at 256 antennas in the reported simulations.","With more than six RF chains, the proposed hybrid precoder reportedly exceeds the spectral efficiency of a fully digital WMMSE system using the same antenna array.","The imperfect-CSI variant converts a worst-case max-min problem into geometric-program form and retains most of the spectral efficiency at channel-error levels up to 30 percent of the channel gain.","Because the loop only depends on gradients, projections, and power normalization, the framework is stated to generalize to MU-MIMO and other hybrid architectures."],"supporting_citations":[{"why":"Supplies the manifold optimization alternating-minimization baseline that GGML is compared against and claims to outperform.","marker":"[9]"},{"why":"Provides the Element-AltMax two-stage hybrid precoding baseline and the RF-chain condition for reaching the fully digital upper bound.","marker":"[8]"},{"why":"Provides the attention-based CNN hybrid precoding baseline that GGML competes with as a data-driven method.","marker":"[28]"},{"why":"Supplies the majorization-minimization alternating-minimization baseline used in the convergence and spectral-efficiency comparisons.","marker":"[43]"},{"why":"Establishes the gradient-based manifold meta-learning approach that GGML extends to hybrid precoding.","marker":"[35]"},{"why":"Supplies the metalearning-based alternating minimization framework for nonconvex optimization that motivates the local-global update structure.","marker":"[36]"},{"why":"Introduces the Kolmogorov-Arnold network architecture that GGML embeds as its update generator.","marker":"[40]"},{"why":"Provides the WMMSE algorithm that GGML is claimed to surpass when the number of RF chains exceeds six.","marker":"[15]"},{"why":"Provides the worst-case robust design framework used to transform the imperfect-CSI maximin problem into a tractable effective-channel problem.","marker":"[44]"}],"fun_headline_variants":["Meta-learning hybrid precoder beats fully digital WMMSE","Gradient-guided meta-learning speeds up hybrid precoding 8x","Online meta-learning solution surpasses fully digital WMMSE","No pretraining: meta-learned hybrid precoding outdoes WMMSE","Hybrid precoding via meta-learning converges 8x faster"],"cache_read_input_tokens":20224,"weakest_assumption_plain":"The method assumes that a randomly initialized small network, trained online for a fixed 500 iterations on a single channel realization, will always produce productive update directions and converge to a good precoder, for any channel and any system size.","fun_headline_variants_meta":{"raw":{"variants":["Meta-learning hybrid precoder beats fully digital WMMSE","Gradient-guided meta-learning speeds up hybrid precoding 8x","Online meta-learning solution surpasses fully digital WMMSE","No pretraining: meta-learned hybrid precoding outdoes WMMSE","Hybrid precoding via meta-learning converges 8x faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000576,"raw_usage":{"total_tokens":2720,"prompt_tokens":951,"completion_tokens":1769,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":1681}},"tokens_in":567,"tokens_out":1769,"duration_ms":12662,"temperature":1.0,"reasoning_tokens":1681,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:56:29.169518+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GGML on a larger or ill-conditioned system than those tested, such as N = 512 or K = 16 with near-degenerate scattering clusters, and check whether a fresh random seed still reaches the claimed spectral efficiency within 500 iterations; if some seeds stall below the manifold-optimization baseline or diverge, the free-of-training claim fails.","supporting_citations":[{"cited_title":"Alternati ng minimization algorithms for hybrid precoding in millimeter wave mimo sys tems,","cited_arxiv_id":null,"evidence_quote":"Supplies the manifold optimization alternating-minimization baseline that GGML is compared against and claims to outperform."},{"cited_title":"Atten tion- based hybrid precoding for mmwave mimo systems,","cited_arxiv_id":null,"evidence_quote":"Provides the attention-based CNN hybrid precoding baseline that GGML competes with as a data-driven method."},{"cited_title":"Hybrid beamforming for mil limeter wave multi-user mimo systems using learning machine,","cited_arxiv_id":null,"evidence_quote":"Supplies the majorization-minimization alternating-minimization baseline used in the convergence and spectral-efficiency comparisons."},{"cited_title":"Robust beamforming for ris - aided communications: Gradient-based manifold meta learn ing,","cited_arxiv_id":null,"evidence_quote":"Establishes the gradient-based manifold meta-learning approach that GGML extends to hybrid precoding."},{"cited_title":"Metalearning-based alternating minimization algorithm for nonconvex optimization,","cited_arxiv_id":null,"evidence_quote":"Supplies the metalearning-based alternating minimization framework for nonconvex optimization that motivates the local-global update structure."},{"cited_title":"An Iterativ ely Weighted MMSE Approach to Distributed Sum-Utility Maximization for a MIMO Interfering Broadcast Channel,","cited_arxiv_id":null,"evidence_quote":"Provides the WMMSE algorithm that GGML is claimed to surpass when the number of RF chains exceeds six."},{"cited_title":"Robust sum rate optimization for the downlink multiuser mimo systems: Wors t-case design,","cited_arxiv_id":null,"evidence_quote":"Provides the worst-case robust design framework used to transform the imperfect-CSI maximin problem into a tractable effective-channel problem."}],"review_version":1}