{"id":"52df9c3c-6ec0-434e-8f21-2b991c8fda5a","arxiv_id":"2608.12918","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"OTA channel labels derived from noisy received signals carry input-correlated supervision noise, so learned channel estimators peak at intermediate measured-data fractions, while end-to-end receivers with error-free bit supervision improve monotonically.","lead":"This paper measures what happens when learned wireless receivers are trained on a mix of simulated and real over-the-air signals. It finds that channel estimators get worse when too much real data is used, because their training labels carry noise correlated with their input, while end-to-end receivers keep improving.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Sec. V-C denoising experiment does not isolate input-correlated supervision noise from label-noise variance or finite-data overfitting, leaving the causal mechanism not uniquely confirmed.","rationale":"I agree with the reader's acceptance but not with the choice of weakest assumption. The CRC filtering assumption is likely benign: with a rate-0.25 code and fewer than 1% CRC failures, undetected errors are rare, and the paper discloses the selection bias. The capacity-matching issue is acknowledged, and the ACM-E degradation does not depend on the CM DeepRx comparison. The load-bearing issue is that the empirical validation of the central mechanism does not isolate the correlation component of the label noise. Equation (6) proves a bias in the infinite-data limit, but the experiments at pm=100 use only about 44k OTA slots with repeated epochs, so the observed degradation could in principle be finite-data overfitting. The denoising experiment changes the target in a way that reduces variance and acts as a regularizer, so it supports 'label quality matters' but not uniquely 'correlation matters.' A synthetic correlated-versus-independent label-noise comparison would settle this. Because the theoretical argument is sound and the denoising result is in the right direction, this is a limitation rather than a fatal flaw; ACCEPT remains appropriate, though the paper would be stronger with the isolation test.","tokens_in":7788,"tokens_out":10317,"duration_ms":118770,"concrete_test":"Run a Sionna-based synthetic experiment with known H. Condition A (correlated): generate Y = Hx + n and label H_gt = Y/x (same n). Condition B (independent): generate Y = Hx + n1 and label H_gt = H + n2/x with n2 independent of n1 at matched variance; use effectively unlimited samples and the same ACM-E training. If Condition A reproduces the high-OTA degradation or an NMSE error floor while Condition B does not, input-correlation is confirmed. If both degrade similarly, or if performance improves with dataset size in Condition A, the current evidence is compatible with label-noise variance or overfitting, and the central claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's causal claim—that the high-OTA degradation of ACM-E is caused specifically by input-correlated supervision noise, not by limited data diversity—rests on the denoising experiment in Sec. V-C. That experiment smooths the OTA label y/x with the same kernel as DDCE and shows improvement at matched OTA fractions. But smoothing changes three things at once: it reduces the variance of label noise (whether correlated or independent), it introduces a bias relative to the true channel, and it acts as a regularizer against overfitting to the finite ~44k-slot OTA dataset. Because the OTA fraction is varied by replacing synthetic data with a fixed pool of measured slots, high pm also means more repeated exposure to the same finite OTA samples and less regularization from online synthetic data. The observed peak at intermediate pm and the improvement after denoising are therefore also consistent with a finite-data/overfitting explanation. Equation (6) proves that in the limit of infinite data the correlated label noise leaves a bias, but the experiments never approach that limit; they do not show that more OTA data at pm=100 fails to close the gap. The 'controlled denoising' manipulation does not isolate the correlation component: applying the same smoothing to labels corrupted by independent noise would also improve performance. Thus the specific mechanism named in the title and abstract is not uniquely identified by the reported experiments.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a 5.88 GHz over-the-air (OTA) OFDM measurement campaign and uses it to train two learned receivers with controlled mixtures of measured and synthetic data: ACM-E, a neural channel estimator, and CM DeepRx, an end-to-end LLR receiver. The central empirical finding is an asymmetry: CM DeepRx improves monotonically as the OTA fraction increases, while ACM-E peaks at an intermediate OTA fraction and degrades at p_m = 100%. The paper attributes this to input-correlated supervision noise: OTA channel labels are constructed as y_ij/x_ij, so the label error shares the noise realization present in the network input, and the MSE objective converges to a conditional expectation that retains a bias (Eq. 6). A denoising experiment that smooths the OTA labels with the DDCE kernel is presented as confirming the mechanism. The paper concludes with practical guidance: end-to-end receivers benefit from fully measured training, whereas channel estimators require improved supervision quality rather than simply more OTA data.","tokens_in":8050,"tokens_out":4309,"duration_ms":45352,"significance":"If the causal claim holds, the paper gives the first systematic OTA evidence that the value of measured training data depends on the supervision target, and it identifies a mechanism—input-correlated supervision noise—that is distinct from the usual independent-label-noise setting. The measurement campaign is substantial, the experiments use five seeds with standard deviations on the headline metric, and the evaluation is on held-out OTA data, which is a clear strength. The theoretical argument in Sec. IV is a straightforward and correct application of the conditional-expectation property of MSE training, and the within-ACM-E denoising intervention changes only the target construction. These strengths make the paper a useful contribution to the practical literature on training learned receivers with real-world data, provided the causal attribution is supported or appropriately qualified.","major_comments":[{"comment":"The denoising experiment does not uniquely identify input-correlated supervision noise as the cause of the high-OTA degradation. Smoothing the label y_ij/x_ij with the DDCE kernel simultaneously (i) reduces the variance of the label noise regardless of whether that noise is correlated with the input, (ii) introduces a bias relative to the true channel, and (iii) acts as a regularizer against overfitting to the fixed pool of roughly 44,000 measured slots, which is increasingly re-used as p_m grows. Equation (6) proves that in the infinite-data limit correlated label noise leaves a residual bias, but the experiments do not approach that limit and do not show that more unique OTA data at p_m = 100% fails to close the gap. A control that corrupts synthetic labels with independent noise of the same variance, or one that compares smoothed correlated versus smoothed independent noise at matched variance, is needed to support the abstract's causal claim that the correlation, rather than limited data diversity or finite-data overfitting, caused the degradation.","section":"Sec. V-C"},{"comment":"The claim that CRC-validated decoded bits provide 'effectively error-free' supervision rests on an unquantified assumption: the paper discards the fewer than 1% of slots failing CRC, but it does not bound the rate of undetected decoding errors in accepted frames. If residual undetected errors are non-negligible, CM DeepRx's bit targets would also carry input-correlated noise, and the asymmetry attributed to the supervision objective would be confounded. Please provide an estimate of the residual post-CRC frame error probability, for example by re-encoding and re-modulating decoded bits and comparing with the received signal on held-out data, or explain quantitatively why code rate 0.25 combined with CRC makes this assumption safe.","section":"Sec. III-D"},{"comment":"The statement that comparing architectures at similar parameter counts 'isolates the benefit of end-to-end optimization from differences in model size' is stronger than the evidence supports. CM DeepRx has 475K parameters versus 412K for ACM-E (about 15% more), roughly 1.9 times the MFLOPs, and a different architecture family and inductive bias. The cross-architecture asymmetry could therefore be influenced by architecture rather than purely by the supervision objective. The within-ACM-E denoising experiment does control architecture, but the headline comparison of 'capacity-matched' models should either be softened or supplemented by a sensitivity analysis over parameter count or latent dimension.","section":"Sec. III-C"}],"minor_comments":[{"comment":"At the headline operating point, the raw-OTA degradation of ACM-E is small: the required SINR at BER = 10^-2 is 16.07 dB at p_m = 100% versus 15.86 dB at p_m = 0%, a 0.21 dB gap. This is significant relative to the standard deviations reported, but the practical magnitude should be stated more explicitly so that readers can calibrate the 'degradation' claim.","section":"Sec. V-B / Table IV"},{"comment":"The notation 'E[H_ij | Y] + E[n_ij/x_ij | Y]' in Eq. (6) is correct, but the sentence 'the label error shares the same realization of measurement noise as the input' could be misread as saying the error is exactly n_ij; it is n_ij/x_ij, which is known only up to the transmitted symbol. A brief clarifying remark that x_ij is known at training time would improve readability.","section":"Sec. IV"},{"comment":"The phrase 'controlled denoising experiment' is potentially misleading because the intervention changes the variance and bias of the target in addition to the correlation structure; consider renaming it a 'target-smoothing experiment' or explicitly listing the three confounded effects.","section":"Sec. V-C"},{"comment":"Reference [12] is typeset as 'AI - Based' with an unintended space; this should be corrected.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The empirical study is valuable and the central asymmetry is clearly demonstrated, but the causal attribution to input-correlated supervision noise is the load-bearing claim of the paper, and the current denoising experiment does not separate that mechanism from label-noise variance reduction or finite-data regularization. I would be willing to accept after the authors either add an appropriate control experiment (e.g., matched-variance independent label noise) or substantially qualify the causal language in the abstract and conclusions. The reliance on the authors' earlier OTA validation paper [12] is appropriate and not excessive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper reports a genuine, useful empirical finding. On a 5.88 GHz measurement campaign, an end-to-end learned receiver keeps improving as you add OTA data, while a learned channel estimator peaks around 40% OTA and degrades at 100%. That asymmetry is new, and the proposed explanation — OTA channel labels are built from the same noisy received signal the network sees as input, so the label noise is correlated with the input and cannot be averaged away — is correct as a theoretical statement. The authors also run a label-denoising experiment intended as a causal check.\n\nWhat the paper does well: the dataset is substantial (~44k OFDM slots across urban, suburban, rural, and industrial environments, pedestrian and vehicular mobility), the training protocol is careful (five seeds, standard deviations on the headline metrics, controlled OTA fractions), and the write-up is honest about limitations. The MSE conditional-expectation argument in Eq. (6) is textbook but applied cleanly to a practical problem. The practical guidance — use fully measured data for end-to-end receivers, but improve label quality for intermediate estimators — is actionable and clearly stated.\n\nThe soft spots. The denoising experiment in Sec. V-C does not isolate the correlation component. Smoothing the labels changes three things at once: it lowers label-noise variance, it biases the target relative to the genie channel, and it acts as a regularizer against overfitting to the finite ~40k-slot OTA training set. The improvement after denoising is consistent with the correlation story, but also with a finite-data/overfitting story. Because pm=100 means roughly 80 passes over the same OTA slots (with input augmentation), high-pm degradation could be partly memorization. The paper's claim that more measurements cannot remove the bias is true in the infinite-data limit, but the experiments never approach that limit; they don't show that more OTA data at pm=100 fails to close the gap. That is the main gap.\n\nA secondary caveat: the capacity matching is approximate. CM DeepRx has about 15% more parameters and 50% more FLOPs, and is trained four times longer, so the end-to-end versus modular comparison is not purely about supervision. The authors acknowledge this, so it is a minor issue, not a fatal one.\n\nOverall, the empirical finding is solid and the theoretical framing is sensible, but the causal claim — that correlation, not data diversity, causes the degradation — is overstrong relative to the evidence. A serious referee should ask for a control with independent label noise (e.g., corrupting targets with synthetic noise uncorrelated with the input) or a finite-data analysis that separates correlation from overfitting. That is a major-revision request, not a rejection; the core contribution survives.\n\nRecommendation: send it to peer review. Anyone working on learned receivers or OTA training will want this paper, and the authors have the measurements and careful methodology to address the remaining questions.","headline":"A careful OTA measurement study showing a real asymmetry between end-to-end and channel-estimator receivers as OTA fraction increases, with a plausible but not fully pinned-down causal mechanism.","tokens_in":784,"tokens_out":1834,"would_cite":true,"duration_ms":47004,"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":"The paper shows that increasing over-the-air training data monotonically improves an end-to-end neural receiver, but degrades a neural channel estimator past an optimum because its labels inherit noise from the input.","keywords":["over-the-air training","neural receivers","channel estimation","supervision noise","OFDM","end-to-end learning","label noise","measurement campaign"],"falsifier":"A decisive test: in a simulator with unlimited samples, corrupt synthetic channel labels as $H + \\alpha n/x$ with the same noise realization $n$ that appears in the input, and sweep $\\alpha$ from 0 (independent noise) to 1 (fully correlated). If the fully correlated case reproduces the channel estimator's peak-and-degrade pattern at high OTA fractions and the independent case does not, the paper's mechanism is confirmed; if both degrade equally, the cause is elsewhere.","tokens_in":7581,"feed_emoji":"📡","tokens_out":8302,"duration_ms":74167,"temperature":0.7,"pith_summary":"The paper asks whether measured over-the-air (OTA) data actually helps learned wireless receivers and shows the answer depends on what the network is trained to predict. Training a capacity-matched end-to-end receiver and a neural channel estimator on mixtures of simulated and measured 5.88 GHz OFDM data, it finds that the end-to-end receiver improves monotonically as the measured fraction grows, while the channel estimator peaks at an intermediate fraction and degrades under fully measured training. The cause is the supervision target: OTA channel labels are built by dividing the received signal by the known transmitted symbol, so the label noise is the same noise realization present at the network input. An MSE-trained estimator converges to the observation-derived target rather than the clean channel, and additional measurements reduce variance but cannot remove the residual bias. CRC-validated decoded bits, by contrast, provide effectively error-free supervision, which is why end-to-end training keeps gaining from measured data.","feed_headline":"Measured data lifts end-to-end receivers, caps channel estimators","feed_subtitle":"Channel labels built from noisy received signals carry input-tied errors; more data can't fix them.","key_machinery":"The load-bearing object is the OTA channel-label construction $H^{\\mathrm{gt}}_{ij}=y_{ij}/x_{ij}$, which extends the least-squares pilot estimate to the full time-frequency grid. Its error term $n_{ij}/x_{ij}$ shares the noise realization with the network input $Y$, so under MSE training the estimator converges to the conditional expectation of the label given the input and retains a bias that extra data cannot remove; this is what the paper calls input-correlated supervision noise. The controlled intervention is the same separable time-frequency average used by decision-directed estimation (kernels $k_f=5$, $k_t=3$), applied to the labels only while input, architecture, and optimization stay fixed. That isolates the supervision mechanism from measurement diversity.","core_discovery":"The central claim is that OTA training data is not uniformly beneficial: it consistently helps an end-to-end receiver (CM DeepRx) while limiting a capacity-matched channel estimator (ACM-E) at high measured-data fractions. The paper attributes this asymmetry to input-correlated supervision noise. At each resource element, the OTA channel label is $H^{\\mathrm{gt}}_{ij}=y_{ij}/x_{ij}=H_{ij}+n_{ij}/x_{ij}$, where $n_{ij}$ is the same noise realization already present in the input $Y$. Because the MSE optimum is the conditional expectation of the target given the input, the correlated error term contributes a non-zero bias that does not vanish as the dataset grows. Bit-level supervision from CRC-validated decoded bits does not inherit input noise, so the end-to-end receiver keeps improving with OTA fraction. A controlled experiment that denoised the channel labels with separable time-frequency smoothing removed most of the high-OTA degradation, confirming that input-correlated supervision noise, not limited data diversity, is responsible.","pith_inferences":["The same mechanism should bind any learned intermediate estimator whose target is constructed from the received signal, such as MIMO channel estimators using pilot-division labels; the paper notes MIMO is out of scope, but the argument does not rely on SISO-specific structure beyond the label construction.","A testable extension: replacing noisy OTA channel labels with estimates from a separate clean reference pass, such as repeated pilot averaging or a second receiver chain, should push the optimal OTA fraction toward 100%, because the bias term would then be uncorrelated with the input.","The paper's asymmetry suggests a design heuristic for future receiver pipelines: prefer bit-level or semantic supervision over intermediate physical quantities whenever the intermediate quantity is observed through the same noisy channel; alternatively, learn the intermediate quantity with a self-supervised objective that avoids noisy explicit targets.","The optimal OTA fraction for channel estimators is likely deployment-dependent, but the qualitative shape of improvement followed by decline should transfer across architectures and frequencies; practitioners can probe it with a small denoising sweep rather than assuming more measurements are always better."],"forward_implications":["End-to-end neural receivers trained on measured data should be trained with the largest available OTA fraction; the paper observed monotonic improvement through $p_m=100\\%$.","Neural channel estimators should be trained with moderate OTA fractions (about 40% in this study) and denoised labels; fully measured raw labels can underperform simulation-only training.","Adding more OTA measurements cannot cure bias caused by input-correlated supervision noise; the fix is to change the target, not grow the dataset.","When supervision targets must be derived from noisy observations, label denoising should be applied before training, for example the same smoothing used by decision-directed estimation.","The same principle constrains any learned receiver stage whose target is a function of the received signal, not just channel estimators in OFDM."],"supporting_citations":[{"why":"Supplies the DeepRx end-to-end receiver architecture that the paper's CM DeepRx adapts with a ConvMixer backbone.","marker":"[3]"},{"why":"Provides pilot-arrangement channel estimation techniques whose least-squares division defines the form of the OTA supervision labels.","marker":"[7]"},{"why":"Compares pilot-aided channel estimation methods, providing the conventional baseline context for label construction and interpolation.","marker":"[8]"},{"why":"Introduces the iterative decision-directed channel estimation whose divide-and-smooth steps supply the denoising kernel used in the label-denoising experiment.","marker":"[9]"},{"why":"Supplies the comparative study of iterative channel estimators underlying the DDCE reference scheme.","marker":"[10]"},{"why":"Defines the Urban Macro channel model used to generate the synthetic training data.","marker":"[17]"},{"why":"Supplies the simulator used for online synthetic data generation during mixed training.","marker":"[18]"},{"why":"Frames the correlated label-noise setting that input-correlated supervision noise is identified as an instance of.","marker":"[19]"}],"fun_headline_variants":["OTA training aids end-to-end receivers, limits channel estimators","Input-correlated noise blunts OTA gains for channel estimators","OTA data helps end-to-end, hurts channel estimators","Correlated supervision noise caps OTA channel estimator gains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, stated in Sec. III-D, is that CRC-validated decoded bits are effectively error-free supervision: residual undetected decoding errors are assumed negligible, so the end-to-end bit targets do not inherit input-correlated noise.","fun_headline_variants_meta":{"raw":{"variants":["OTA training aids end-to-end receivers, limits channel estimators","Input-correlated noise blunts OTA gains for channel estimators","OTA data helps end-to-end, hurts channel estimators","Correlated supervision noise caps OTA channel estimator gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000419,"raw_usage":{"total_tokens":2171,"prompt_tokens":975,"completion_tokens":1196,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":1130}},"tokens_in":591,"tokens_out":1196,"duration_ms":9222,"temperature":1.0,"reasoning_tokens":1130,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:37:07.969911+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test: in a simulator with unlimited samples, corrupt synthetic channel labels as $H + \\alpha n/x$ with the same noise realization $n$ that appears in the input, and sweep $\\alpha$ from 0 (independent noise) to 1 (fully correlated). If the fully correlated case reproduces the channel estimator's peak-and-degrade pattern at high OTA fractions and the independent case does not, the paper's mechanism is confirmed; if both degrade equally, the cause is elsewhere.","supporting_citations":[{"cited_title":"DeepRx: Fully Convolutional Deep Learning Receiver,","cited_arxiv_id":null,"evidence_quote":"Supplies the DeepRx end-to-end receiver architecture that the paper's CM DeepRx adapts with a ConvMixer backbone."},{"cited_title":"Channel estimation techniques based on pilot arrangement in OFDM systems,","cited_arxiv_id":null,"evidence_quote":"Provides pilot-arrangement channel estimation techniques whose least-squares division defines the form of the OTA supervision labels."},{"cited_title":"A comparison of pilot-aided channel estimation methods for OFDM systems,","cited_arxiv_id":null,"evidence_quote":"Compares pilot-aided channel estimation methods, providing the conventional baseline context for label construction and interpolation."},{"cited_title":"CD3-OFDM: a novel demodulation scheme for fixed and mobile receivers,","cited_arxiv_id":null,"evidence_quote":"Introduces the iterative decision-directed channel estimation whose divide-and-smooth steps supply the denoising kernel used in the label-denoising experiment."},{"cited_title":"A comparative study of iterative channel estimators for mobile ofdm systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the comparative study of iterative channel estimators underlying the DDCE reference scheme."},{"cited_title":"Study on channel model for frequencies from 0.5 to 100 GHz,","cited_arxiv_id":null,"evidence_quote":"Defines the Urban Macro channel model used to generate the synthetic training data."},{"cited_title":"Classification in the Presence of Label Noise: A Survey,","cited_arxiv_id":null,"evidence_quote":"Frames the correlated label-noise setting that input-correlated supervision noise is identified as an instance of."}],"review_version":1}