{"id":"89c84808-10bf-438e-8849-bfb127ce2e5d","arxiv_id":"2606.06723","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"CNN learns sparse array designs for single-source single-interferer beamforming via pre-steering, reporting over 90% test accuracy across angles.","lead":"This paper trains a convolutional neural network to pick sparse antenna array layouts that form beams picking up one desired signal while rejecting one interferer. A pre-steering step lets the network train only at broadside and then handle other angles without retraining, plus extra training with small steering errors for robustness.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Simulation-to-real gap: single-source/single-interferer synthetic data may not generalize to realistic propagation","rationale":"The reader's weakest assumption directly identifies the load-bearing point. Because the full text (as described) remains confined to the same single-source/single-interferer simulation regime, the concern is unchanged and the unverdicted status is appropriate pending the proposed check.","tokens_in":1699,"tokens_out":310,"duration_ms":12798,"concrete_test":"Generate a new test set with two interferers plus Rayleigh multipath (K=4 taps) at the same angle ranges; apply the trained CNN with pre-steering and measure both classification accuracy and achieved SINR relative to the single-interferer case. If accuracy falls below 75% or SINR loss exceeds 3 dB, the representativeness assumption fails.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim requires that >90% classification accuracy on pre-steered inputs in single-source/single-interferer simulations (with only angular uncertainty modeled via error-augmented training) implies near-optimal SINR and rapid reconfiguration in dynamic environments. This holds only if the synthetic data generation (fixed/varying source angles, pre-steering to broadside, additive noise) captures the dominant effects of real channels. The abstract and described approach give no indication of multi-interferer, multipath, or measured data; all reported accuracy is on the same controlled simulation distribution used for training.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes using convolutional neural networks to select sparse array configurations for adaptive beamforming. A pre-steering strategy aligns test inputs to broadside so that a single network trained at broadside can handle varying source angles; error-augmented training is added for robustness to pre-steering imperfections. Evaluation is performed exclusively on simulated single-source, single-interferer scenarios, with the central empirical claim being that the method achieves over 90% test accuracy and near-optimal SINR across wide angular ranges.","tokens_in":1827,"tokens_out":524,"duration_ms":14449,"significance":"If the accuracy claim were shown to correspond to measurable SINR gains over standard sparse-array design methods and if the simulation distribution were demonstrated to be representative, the pre-steering technique could reduce the need for repeated optimization or retraining in time-varying environments. The current manuscript supplies no such comparisons or generalization evidence.","major_comments":[{"comment":"Abstract: the claim that the method 'achieves near-optimal beamforming' and 'maximizes the SINR' is unsupported because no baseline (random selection, convex optimization, or conventional sparse-array algorithms) is reported and no quantitative definition of 'near-optimal' or SINR computation procedure is supplied.","section":"Abstract"},{"comment":"Abstract and results description: the reported >90% test accuracy is given without error bars, confidence intervals, or statistical tests, and without stating the number of Monte-Carlo trials or the precise train/test split, rendering the accuracy figure impossible to interpret as evidence of reliable performance.","section":"Abstract"},{"comment":"Abstract: the evaluation is restricted to single-source/single-interferer synthetic data with only angular uncertainty; this setup does not address multi-interferer, multipath, or measured-channel conditions that are central to the stated goal of 'highly dynamic propagation environments,' so the generalization claim rests on an untested assumption.","section":"Abstract"}],"minor_comments":[{"comment":"Notation for array geometry, steering vectors, and the precise CNN architecture (layer counts, filter sizes, output classes) should be defined explicitly in the methods section rather than left implicit.","section":null},{"comment":"The manuscript should include a clear statement of the loss function used for training and the exact mapping from network output to array configuration.","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback on our manuscript. We address each major comment point by point below and will revise the manuscript to improve clarity and support for the claims where feasible.","responses":[{"response":"We agree that the abstract makes claims of near-optimal performance and SINR maximization without direct baseline comparisons or explicit definitions. The manuscript positions the CNN approach as an alternative to conventional and convex optimization methods but does not report quantitative comparisons. In the revision we will add a dedicated results subsection with comparisons to random selection, convex optimization, and conventional sparse-array algorithms, along with a precise definition of near-optimality (e.g., SINR within X dB of the optimal configuration) and a clear description of the SINR computation procedure.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the claim that the method 'achieves near-optimal beamforming' and 'maximizes the SINR' is unsupported because no baseline (random selection, convex optimization, or conventional sparse-array algorithms) is reported and no quantitative definition of 'near-optimal' or SINR computation procedure is supplied."},{"response":"We acknowledge that the reported accuracy figure lacks supporting statistical details. The manuscript states 'over 90% test accuracy' without error bars, trial counts, or split information. We will revise the abstract and results sections to specify the number of Monte-Carlo trials, the exact train/test split, and to include error bars or confidence intervals for the accuracy metrics.","revision_made":"yes","referee_comment":"[Abstract] Abstract and results description: the reported >90% test accuracy is given without error bars, confidence intervals, or statistical tests, and without stating the number of Monte-Carlo trials or the precise train/test split, rendering the accuracy figure impossible to interpret as evidence of reliable performance."},{"response":"The work deliberately focuses on single-source/single-interferer scenarios with angular uncertainty to isolate and validate the pre-steering and error-augmented training technique. We agree this does not cover multi-interferer, multipath, or measured channels. We will revise the abstract and discussion to explicitly limit the scope of the claims, remove over-generalization language, and identify multi-interferer and real-channel extensions as future work. No additional simulations will be performed for this revision.","revision_made":"partial","referee_comment":"[Abstract] Abstract: the evaluation is restricted to single-source/single-interferer synthetic data with only angular uncertainty; this setup does not address multi-interferer, multipath, or measured-channel conditions that are central to the stated goal of 'highly dynamic propagation environments,' so the generalization claim rests on an untested assumption."}],"tokens_in":1395,"tokens_out":589,"duration_ms":18167,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper trains a CNN to pick sparse array configurations after pre-steering inputs to broadside, with error-augmented training to handle small steering mistakes. It reports over 90% test accuracy across angle ranges for one desired source and one interferer, both fixed and varying source cases.\n\nThe pre-steering step is the clearest practical move. Training once at broadside and then rotating test inputs avoids retraining the network for every source direction, which matters if the array has to reconfigure quickly. Adding structured perturbations during training is a straightforward way to build some robustness to angle uncertainty.\n\nThe evaluation stays limited. The abstract gives no baselines against convex optimization or greedy selection methods, no error bars on the accuracy numbers, and no breakdown of how the final SINR is actually computed or compared. All results come from the same synthetic single-source single-interferer model used for training, with no multipath, no measured data, and no multiple interferers. That leaves the simulation-to-real question open.\n\nThis is for people already working on learned array configuration who want to see the pre-steering wrapper applied to the standard narrow scenario. It is not a broad advance in beamforming theory or hardware.\n\nI would send it to peer review so the full methods, any hidden comparisons, and the exact data generation can be checked. The core idea is clear enough to be worth that step, even if the current claims rest on narrow evidence.","headline":"CNN plus pre-steering for single-source single-interferer sparse array selection; the evaluation is thin on baselines and stays inside simulation.","tokens_in":2324,"tokens_out":367,"would_cite":false,"duration_ms":12413,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A convolutional neural network trained only at broadside can select sparse array configurations for near-optimal beamforming at arbitrary angles using pre-steering.","keywords":["deep learning","sparse array design","adaptive beamforming","pre-steering","convolutional neural networks","array configuration","signal to interference ratio"],"falsifier":"A hardware experiment measuring actual SINR in a real propagation environment with varying angles, where performance falls well below the simulated levels predicted by the 90% accuracy.","tokens_in":2599,"feed_emoji":"📡","tokens_out":569,"duration_ms":21484,"temperature":0.7,"pith_summary":"The paper shows that CNNs can learn to pick sparse array layouts that give good beamforming performance when a desired signal and an interferer come from changing directions. Pre-steering shifts the inputs so the network trains once for broadside and then handles any source angle without retraining. Robust training with small steering errors keeps accuracy high even when the steering is not perfect. If this holds, it offers a fast alternative to slow optimization methods for reconfiguring arrays in changing environments.","feed_headline":"CNN picks sparse arrays for 90% accurate beamforming","feed_subtitle":"Pre-steering strategy trains the network once at broadside to handle any source angle while keeping high SINR.","key_machinery":"Convolutional neural network that classifies optimal sparse array configurations from pre-steered array response inputs.","core_discovery":"The authors demonstrate that a CNN classifier, trained on pre-steered data with added angular perturbations, can identify sparse array configurations that achieve near-optimal SINR with over 90 percent test accuracy across wide ranges of source and interference angles, for both fixed and varying source directions.","pith_inferences":["If the simulation results hold, this could enable real-time adaptive beamforming in practical systems with changing conditions.","Extending the training to include multiple interferers might broaden applicability without major changes to the pre-steering idea.","The pre-steering strategy could be tested on other array processing tasks like direction finding."],"forward_implications":["The method allows rapid array reconfiguration without per-angle retraining.","Accuracy remains high even with pre-steering errors due to error-augmented training.","It supports both fixed and varying desired source directions.","The approach achieves over 90% test accuracy for single source and single interferer scenarios."],"fun_headline_variants":["CNN pre-steers inputs for sparse array selection at 90% accuracy","Single CNN training at broadside covers all source angles with pre-steering","Error-augmented training lets CNN handle pre-steering imperfections","CNN identifies sparse arrays for 90% accurate SINR in varying angles"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"That the simulated single-source single-interferer scenarios with pre-steering represent real propagation environments closely enough for the accuracy to carry over to actual systems.","fun_headline_variants_meta":{"raw":{"variants":["CNN pre-steers inputs for sparse array selection at 90% accuracy","Single CNN training at broadside covers all source angles with pre-steering","Error-augmented training lets CNN handle pre-steering imperfections","CNN identifies sparse arrays for 90% accurate SINR in varying angles"]},"model":"grok-4.3","cost_usd":0.005804,"raw_usage":{"total_tokens":2747,"prompt_tokens":636,"num_sources_used":0,"completion_tokens":75,"cost_in_usd_ticks":58037000,"prompt_tokens_details":{"text_tokens":636,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2036,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":636,"tokens_out":75,"duration_ms":12937,"temperature":1.0,"reasoning_tokens":2036,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T23:49:45.153598+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A hardware experiment measuring actual SINR in a real propagation environment with varying angles, where performance falls well below the simulated levels predicted by the 90% accuracy.","supporting_citations":[],"review_version":1}