{"id":"209e1621-b378-48a3-9f91-902e0dbce670","arxiv_id":"2504.14835","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GFL4BS uses zero-shot synthetic data generation and per-modality federated aggregation to improve beam-selection accuracy under label and modality imbalance.","lead":"Federated learning for vehicle-to-everything beam selection struggles when cars have very different labels or sensor sets. This paper adds a generator that creates synthetic training data from the global model's statistics, reporting accuracy gains over two federated baselines on the FLASH mmWave dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The generator's central assumption—that global-model BN statistics encode the true data distribution—is unvalidated under label imbalance, and the biased evaluation model Me can make synthetic data reinforce, not correct, the imbalance.","rationale":"The reader's weakest assumption identifies exactly the load-bearing concern: the AMD generator's reliance on global-model BN statistics and a biased evaluation model. I agree this is the central risk to the paper's core claim. If synthetic data do not represent the true minority-class distribution, then GFL4BS is not actually correcting label imbalance; it is merely augmenting with samples that match the biased model, which could reinforce the bias. This concern is especially acute because the paper discloses the caveat about low global-model accuracy and because Me is explicitly trained on the imbalanced scenario, creating a self-referential loop. The proposed test would settle the concern by checking whether generated minority samples are recognized by an unbiased classifier and are close to real minority features. I also note a secondary inconsistency—the abstract's '16.2% higher accuracy under severe label imbalance' is actually the original-dataset difference from Table I (88.5% vs 72.3%), while the severe H row shows a 17.4-point difference—but this is a reporting error, not a fundamental flaw. The methodological gap is more load-bearing because it threatens the validity of the proposed mechanism. Since the paper provides no direct evidence on synthetic data quality, the conditional verdict is appropriate: the claims are plausible but not fully validated. Therefore, I keep the reader's CONDITIONAL verdict unchanged.","tokens_in":8944,"tokens_out":4343,"duration_ms":43537,"concrete_test":"Train an unbiased reference classifier on the centralized balanced FLASH dataset. Under the paper's H-imbalance setting (ζ=24.7%), use the AMD generator to synthesize minority-label samples. Then measure (a) the reference classifier's Top-1 accuracy on these synthetic samples and (b) the feature-space distance (e.g., Fréchet Inception Distance on intermediate features) between synthetic and real minority samples. If the reference accuracy is near chance or the feature distance is large relative to majority-class synthetic data, the BN-statistic generation does not represent the true minority distribution, and the claimed bias-correction mechanism is unsupported. As a complementary check, compare GFL4BS against a trivial oversampling baseline that duplicates real minority samples; if GFL4BS does not outperform it, the generative component is not contributing distributional information.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of GFL4BS is the AMD generator, whose loss in Eq. (5) optimizes synthetic data to match the BN running means and variances of the global model while satisfying hard labels and soft labels from an evaluation model Me. For this to correct label imbalance, the BN statistics must encode enough of the true data distribution that the generated minority-label samples are actually representative of the missing classes. This is not established. Under severe imbalance, the global model is trained on skewed data, so its BN statistics are dominated by majority classes; the cross-entropy terms H(ˆchard, ˆcp) and H(˜c, ˆcp) then force the generator to produce samples that the biased global model and the equally biased Me (obtained by FedAvg on the same imbalanced data, as stated in the paper) classify as the minority label. Such samples may lie in low-density or adversarial regions rather than the true p(x|y) for that class. The paper itself notes that \"if the global model exhibits low accuracy, the quality of the generated data may be compromised,\" but it does not provide any validation of synthetic data quality, such as visual inspection, feature-space comparison with real data, or an independent classifier. Without that validation, the claimed \"correction\" of label imbalance is not supported; the observed accuracy gains could come from sample-count balancing or from fitting the biased decision boundary, rather than from genuinely recovering the missing data distribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GFL4BS, a generative federated learning framework for beam selection in V2X systems. It addresses label and modality imbalances by having each vehicle synthesize multi-modal training data with an AMD generator that matches the batch-normalization statistics of the global model under hard and soft label constraints (Eq. (5)), binarizes synthetic LiDAR, and then performs distributed local training with branch-wise aggregation at the BS. Experiments on the FLASH dataset compare GFL4BS with FedAvg, FLASH, and centralized learning under varying label imbalance and modality loss, reporting higher Top-1 accuracy and sum-rate ratios. The code is publicly available.","tokens_in":9231,"tokens_out":6277,"duration_ms":54659,"significance":"If the reported gains are robust, GFL4BS would be a useful contribution to privacy-preserving, sensing-aided beam alignment with heterogeneous sensors. The paper has concrete strengths: evaluation on a public dataset, a clear system model, and a communication-overhead analysis. However, the central mechanism is not yet convincingly validated: the generator relies on BN statistics of biased models, the reported gains rest on single-run point estimates, and key components lack ablations. The headline claim is also inconsistent with the table. These issues are fixable but require additional experiments and analysis.","major_comments":[{"comment":"The abstract states that GFL4BS achieves '16.2% higher accuracy than the current state-of-the-art under severe label imbalance conditions,' but the 16.2% figure in Table I is the Top-1 accuracy gap between GFL4BS and FLASH on the Original (ζ=78.9%) dataset, not the severe-imbalance row. In the High-imbalance (ζ=24.7%) row the absolute Top-1 accuracy gap is 74.3−56.9=17.4 points. Please correct the headline to match the table and state whether improvements are absolute percentage points or relative improvements.","section":"Abstract and §IV-C, Table I"},{"comment":"The data-generation mechanism is self-referential in a way that is not addressed. Eq. (5) optimizes synthetic inputs to match the BN running statistics of the global model, which is trained on the same imbalanced data, and the soft-label term H(˜c, ˆcp) uses Me, an evaluation model explicitly obtained by FedAvg in the same imbalanced scenario. If both models are biased toward majority classes, the generator may be pushed to synthesize minority-label samples in low-density or adversarial regions rather than in the true class-conditional distribution. The paper acknowledges in §III-A that low global-model accuracy can compromise generated data, but it provides no validation of synthetic data quality, such as visual inspection, feature-space comparison with real data, or an independent classifier. Please add such validation and, ideally, an ablated variant that replaces generated minority samples with real minority samples or simple oversampling to demonstrate that the gains come from distribution recovery rather than count balancing.","section":"§III-A, Eq. (5), and §IV"},{"comment":"All results are single-run point estimates; no standard deviations, confidence intervals, or multiple seeds are reported. Given the reported inter-vehicle variance (e.g., 99.2 for GFL4BS versus 249.1 for FedAvg in the High-imbalance Acc column), the Top-1 accuracy differences in Table I and Fig. 3 may not be statistically significant. Please report means and variances over at least three independent runs, or otherwise establish significance.","section":"§IV-C, Table I and §IV-D, Fig. 3"},{"comment":"No ablation study isolates the contributions of the AMD generator, the soft-label term H(˜c, ˆcp), the BN-statistic matching, the binarization threshold τ, the trigger threshold γ, or the spectral-regularized loss. Without ablations, the causal claim that the generator and spectral-regularized loss 'enhance the expressiveness of synthetic data' is not supported. In particular, a comparison against simple class-balanced re-sampling of real data would clarify whether the generator adds value beyond rebalancing.","section":"§III-A–§III-C, Algorithm 1"},{"comment":"The LiDAR binarization in Eq. (7) maps values to {0,1} by thresholding, but the text immediately states that this step 'embed[s] the receiver and transmitter positions as –2 and –1.' A binary threshold function cannot produce -2 and -1, so the description of how synthetic LiDAR preserves the original integer alphabet {0,1,−1,−2} is internally inconsistent. Please clarify the post-processing that assigns these values or revise Eq. (7).","section":"§III-A, Eq. (7)"},{"comment":"The term 'spectral-regularized loss' is used to describe Eq. (5), but the expression contains no spectral-domain or spectral-norm regularization term; it is a BN-statistic matching loss plus two cross-entropy terms. As the spectral regularity is one of the two 'core innovations' in the abstract, either define the spectral regularization used or rename the loss to avoid an unsupported claim.","section":"Abstract and §III-A, Eq. (5)"}],"minor_comments":[{"comment":"The update step cites Eq. (11a) twice for both the integration and extractor branches; the second citation should presumably be Eq. (11b).","section":"Algorithm 1, line 13"},{"comment":"The evaluation model Me is said to be 'obtained by FedAvg in current imbalanced scenario,' but no architectural details, training epochs, or sharing mechanism are given. Please specify how Me is constructed and used, since it directly influences the soft-label term in Eq. (5).","section":"Footnote 2 and §IV"},{"comment":"The dataset description would benefit from the number of participating vehicles, the local data split details, and a definition of the reported 'Var' column so that readers can interpret the variance numbers in Table I.","section":"§IV-C"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising direction, but the current evidence does not support the headline claims. The most serious issues are the self-referential generation loop with an evaluation model trained on the same imbalanced data, the inconsistency between the abstract's 16.2% claim and the severe-imbalance row, and the lack of any validation of synthetic data quality. The 'spectral-regularized' loss is also not reflected in the equations. These are fixable with additional experiments and careful rewriting, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the combination is new: zero-shot BN-statistics matching from [18] is adapted to multi-modal beam selection, and it is married to FLASH-style per-branch federated aggregation plus a spectral regularizer. Second, treat the headline number with suspicion: the abstract claims 16.2% higher accuracy under severe label imbalance, but that number actually comes from the original, relatively balanced dataset. On the severe imbalance row (Table I) the improvement over FLASH is 17.4 percentage points, which is not the same thing.\n\nWhat is genuinely good: the paper tackles a real deployment problem (label and modality imbalance in federated beam selection for V2X), uses a real mmWave dataset (FLASH), and compares against reasonable baselines. The communication overhead reduction from uploading only the available branches is a practical point that is often ignored. The writing is clear, the method is described in enough detail to reproduce, and the authors released code.\n\nThe soft spots are real but not fatal. The central mechanism—that global-model BN statistics encode enough of the true data distribution to generate useful minority-class samples—is asserted, not validated. The paper admits quality may suffer if the global model is bad, but the proposed fix (Me, a FedAvg model trained on the same imbalanced data) may inherit the same bias, so the soft-label constraint could reinforce rather than correct the imbalance. There is no visual or feature-space check of synthetic data, no independent classifier test, and no ablation isolating the spectral loss, the generator, or the binarization threshold. All numbers are single-run point estimates, so we don't know how stable the gains are. The threshold gamma is said to be insensitive, but no sensitivity analysis is shown.\n\nWho gets value from this: researchers working on federated learning with non-IID or missing-modality data, especially in wireless sensing contexts. It is worth reading as a proof of concept, and it deserves a serious referee—the problem is important, the method is plausible, and the flaws are fixable with additional experiments rather than being conceptual dead ends. I would want error bars, ablations, and some validation of the generated data before accepting the claims.\n\nIn short: give it peer review, but ask for the missing experiments. It is not a breakthrough, but it is honest applied work.","headline":"A competent incremental method paper that combines two known ideas and reports plausible gains, but the abstract's headline number is mismatched and the generative mechanism's core assumption is unvalidated, so the empirical claims need more support before being taken at face value.","tokens_in":9764,"tokens_out":1840,"would_cite":false,"duration_ms":18559,"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":"This paper claims that a generative federated learning method can repair label imbalance and missing sensor modalities in multi-modal beam selection by synthesizing training data from the global model's batch-normalization statistics…","keywords":["beam selection","federated learning","multi-modal sensing","label imbalance","modality imbalance","zero-shot data generation","sensing-aided communication","V2X"],"falsifier":"Take a global model trained only on one beam class, run the AMD generator to create samples for the missing classes, and compare those synthetic samples with held-out real samples from the same classes using a statistical distance or by measuring whether adding them to training improves a balanced test set; if the synthetic samples are statistically unlike the real ones or actually lower accuracy, the central claim fails.","tokens_in":8735,"feed_emoji":"📡","tokens_out":6548,"duration_ms":56161,"temperature":0.7,"pith_summary":"For multi-modal beam selection in vehicle-to-everything systems, the paper tries to establish that federated learning can be made resilient to two practical failures at once: some vehicles see far fewer samples of some beam labels, and others lack whole sensor modalities such as LiDAR or RGB cameras. Its proposal, GFL4BS, has each vehicle generate synthetic multi-modal data from the global model's batch-normalization statistics, with no raw data shared, to rebalance labels and fill missing features before local training. If the claim holds, vehicles with uneven sensors and data can still train a global beam-selection model that approaches centralized-learning accuracy while keeping communication overhead low. The reported results put the Top-1 accuracy gain at 16.2% over the strongest baseline under severe label imbalance, and above-70% success rate even when two vehicles lack both LiDAR and RGB camera inputs.","feed_headline":"Generative FL lifts beam-selection accuracy by 16.2%","feed_subtitle":"Synthetic multi-modal data fills missing labels and sensors, keeping V2X beam links above 70% success.","key_machinery":"The central mechanism is the adaptive zero-shot multi-modal data (AMD) generator. It solves a minimization over synthetic inputs whose loss matches the activations' mean and variance to the batch-normalization running statistics of the global model, adds a hard cross-entropy label constraint and a soft label from an evaluation model, and uses a spectral-regularized variant of this loss when filling missing modalities against the integration branch. A binarization step maps generated LiDAR data back to the discrete voxel set {0,1,-1,-2}. The generator only runs when the global loss decline passes a threshold γ; it is what converts an imbalanced local dataset into a balanced mixed dataset without any client sharing raw data.","core_discovery":"On its own terms, GFL4BS claims that the global model's batch-normalization running means and variances contain enough distribution information that synthetic inputs optimized to reproduce them, under hard- and soft-label constraints, can serve as realistic augmented data for both label balancing and missing-modality completion. This makes data generation zero-shot in the sense of requiring no access to any client's raw data. The paper backs this with simulations on a multi-modal beam-selection dataset: under severe label imbalance GFL4BS reaches 93.3% communication ratio and 74.3% Top-1 accuracy versus 88.0% ratio and 56.9% accuracy for the strongest baseline, and it preserves high accuracy under partial modality loss while transmitting about 25% of the model per round from sensor-poor vehicles.","pith_inferences":["Implicit consequence: the generator's reliance on BN statistics suggests a direct test comparing real and synthetic samples from the same label with distributional metrics; if they diverge, the augmentation will inherit the global model's bias.","One extension: the same BN-statistics-based generation could be repurposed as an inference-time sensor-completion front end, so vehicles with permanently missing LiDAR or RGB could still run the full model rather than only the GPS branch.","Because the method's gains shrink when many vehicles lack entire modalities, a natural next step is to train a small generator on the integration branch's features to model cross-modality correlations before filling large-scale missing inputs."],"forward_implications":["If GFL4BS is correct, federated beam-selection models can retain high Top-1 accuracy under severe label imbalance without any raw data leaving the vehicle, since the synthetic data is derived only from the global model.","Missing RGB or LiDAR features can be reconstructed well enough to recover most of the lost accuracy at partial loss levels, although whole-modality absence across many vehicles remains a hard case.","The per-round communication cost drops for sensor-poor vehicles, which only upload the branches they own, and the synthetic augmentation shortens the number of rounds needed to converge.","The same trigger mechanism generalizes across imbalance levels because the loss-decline threshold γ is insensitive to update step size."],"supporting_citations":[{"why":"Supplies the zero-shot data generation method based on batch-normalization statistics that Eq. (5) adapts for multi-modal data.","marker":"[18]"},{"why":"Supports the assumption that batch-normalization running means and variances reflect the input data distribution.","marker":"[19]"},{"why":"Provides the multi-modal beam-selection dataset and the baseline model that defines the comparison GFL4BS claims to beat by 16.2%.","marker":"[11]"},{"why":"Provides the standard federated-averaging rule and the baseline that GFL4BS compares against for aggregation.","marker":"[20]"},{"why":"Documents the non-IID problem in federated learning and the data-sharing strategy that GFL4BS explicitly avoids.","marker":"[17]"},{"why":"Provides a vertical-federated-learning treatment of modality heterogeneity that GFL4BS positions against.","marker":"[14]"}],"fun_headline_variants":["Zero-shot generative FL fixes label and sensor gaps","GFL4BS: 16.2% better beam selection with missing sensors","Generative FL keeps V2X beam links above 70% success","Hybrid FL training: 16.2% accuracy gain with 25% model upload"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the global model's internal layer statistics faithfully encode what real sensor data looks like, so synthetic samples tuned to match those statistics are genuinely realistic; if the model has already been biased by imbalanced data, the generated data inherits that bias and can reinforce it.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot generative FL fixes label and sensor gaps","GFL4BS: 16.2% better beam selection with missing sensors","Generative FL keeps V2X beam links above 70% success","Hybrid FL training: 16.2% accuracy gain with 25% model upload"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000593,"raw_usage":{"total_tokens":2760,"prompt_tokens":910,"completion_tokens":1850,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":1782}},"tokens_in":526,"tokens_out":1850,"duration_ms":11121,"temperature":1.0,"reasoning_tokens":1782,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:39:22.447992+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a global model trained only on one beam class, run the AMD generator to create samples for the missing classes, and compare those synthetic samples with held-out real samples from the same classes using a statistical distance or by measuring whether adding them to training improves a balanced test set; if the synthetic samples are statistically unlike the real ones or actually lower accuracy, the central claim fails.","supporting_citations":[{"cited_title":"Towards Fair Federated Learning with Zero-Shot Data Augmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the zero-shot data generation method based on batch-normalization statistics that Eq. (5) adapts for multi-modal data."},{"cited_title":"Zeroq: A Novel Zero Shot Quantization Framework,","cited_arxiv_id":null,"evidence_quote":"Supports the assumption that batch-normalization running means and variances reflect the input data distribution."},{"cited_title":"FLASH: F ederated Learning for A utomated S election of H igh-band mmWave Sectors,","cited_arxiv_id":null,"evidence_quote":"Provides the multi-modal beam-selection dataset and the baseline model that defines the comparison GFL4BS claims to beat by 16.2%."},{"cited_title":"Synesthesia of Machines (SoM)-Aided FDD Precoding with Sensing Heterogeneity: A Vertical Federated Learning Approach,","cited_arxiv_id":null,"evidence_quote":"Provides a vertical-federated-learning treatment of modality heterogeneity that GFL4BS positions against."}],"review_version":1}