{"id":"a8493092-6cf3-46f1-ba21-5a0749f86195","arxiv_id":"2607.08717","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":7,"one_line_summary":"Two convolutional neural networks (NBI-CNet and LLR-CNet) jointly estimate and cancel multi-tone narrowband interference in OFDM systems and whiten non-Gaussian residuals for LDPC decoding, eliminating error floors at lower complexity than iterative compressed-sensing baselines.","lead":"This paper proposes two neural networks that jointly cancel narrowband interference in OFDM systems and produce well-calibrated soft bit information for decoding. It matters because it eliminates error floors that plague conventional methods while reducing computational complexity in dense interference scenarios.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"LLR-CNet's learned residual whitening is trained and evaluated exclusively under AWGN with perfect equalization; multipath fading residuals would break the statistical match it relies on.","rationale":"The reader correctly identified the AWGN-only evaluation as the weakest assumption. I agree this is the most load-bearing concern because the paper's central contribution—LLR-CNet as a learned structural whitener—is only as good as the statistical match between training and deployment residuals. Under multipath with imperfect equalization, the residual structure changes fundamentally. However, I rate the verdict as UNCHANGED rather than a downgrade because the paper is transparent about this assumption (§II.C explicitly states it), frames it as isolating the NBI modules, and the contribution of jointly feeding the interference estimate to the LLR estimator remains architecturally sound even if the specific trained weights would need retraining for fading channels. The CONDITIONAL verdict already captures this limitation. The secondary concerns raised by the reader (scale-invariance tested on one FFT size, complexity advantage only at Q>26, train-test d_m mismatch) are valid but less load-bearing: the scale-invariance claim has a physical basis in the Dirichlet kernel scaling (Eq. 40), the complexity crossover is clearly reported, and the d_m mismatch (2 vs 4) actually makes evaluation harder, not easier. The absence of code release is a reproducibility concern but not a correctness concern. The paper's claims are internally consistent and the simulations support them within their stated scope; the scope is simply narrower than the abstract implies.","tokens_in":26085,"tokens_out":762,"duration_ms":254474,"concrete_test":"Retrain NBI-CNet and LLR-CNet under a standardized multipath fading channel model (e.g., 3GPP TDL-A or ETU) with imperfect channel estimation (e.g., pilot-based LS/MMSE with realistic pilot spacing). Evaluate BLER vs. SNR at SIR=-10 dB, Q=8, and compare against the EOMP-IDS + LLR-CNet baseline. If the NBI-CNet + LLR-CNet pipeline exhibits an error floor or the SNR gap to the optimal baseline widens beyond 1 dB, the structural-whitening claim does not generalize to realistic channels.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that LLR-CNet acts as a 'structural whitener' that learns the non-Gaussian statistics of post-mitigation residuals (§VI). This claim depends critically on the residual structure being statistically stable and learnable. However, the entire training and evaluation pipeline assumes H_k=1 (AWGN, §II.C), which the paper justifies as mirroring 'the effective background noise floor of a post-equalization signal.' This assumes perfect equalization. In a real multipath OFDM system, post-equalization residuals include channel estimation error and residual ISI, whose statistics are fundamentally different from AWGN. LLR-CNet takes both the cleaned signal and the estimated interference footprint as inputs (Eq. 34). If the signal component carries residual fading or channel-estimation-error structure that LLR-CNet never saw during training, its learned mapping to 'well-calibrated soft metrics' would be mismatched. The paper does not test this. The AWGN assumption is not merely a simplification—it removes an entire class of structured disturbance that LLR-CNet would need to whiten in practice. This is the soft spot: the 'structural whitener' claim is validated only in the regime where the dominant non-Gaussian structure comes from NBI residuals alone, which is the most favorable case for the proposed method.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The paper proposes a unified deep-learning framework for joint narrowband interference (NBI) cancellation and soft demodulation in OFDM systems. The framework comprises two components: NBI-CNet, a physics-informed convolutional network that estimates multi-tone NBI parameters (gain, fractional frequency offset, phase) in a single forward pass without requiring prior knowledge of the interferer count; and LLR-CNet, a lightweight neural network that maps non-Gaussian post-mitigation residuals onto calibrated log-likelihood ratios (LLRs) for LDPC decoding. The authors show via simulation that the joint architecture eliminates error floors inherent to classical compressed-sensing baselines (OMP-IDS, EOMP-IDS) paired with Gaussian max-log demappers, operates within 0.2–0.5 dB of the optimal iterative baseline under severe interference (SIR=−10 dB), and delivers >3 dB coding gain under mild interference (SIR=10 dB, Q=12) where greedy algorithms suffer signal-peak confusion. The architecture is shown to generalize across FFT sizes without retraining via a closed-form gain rescaling.","tokens_in":26243,"tokens_out":1406,"duration_ms":272222,"significance":"The paper addresses a practically important problem: the mismatch between imperfect NBI cancellation and downstream soft demodulation in OFDM systems. The joint architecture is well-motivated, and the physics-informed design of NBI-CNet—explicitly modeling Dirichlet kernel spectral leakage with a differentiable reconstruction layer—is a strength. The complexity analysis (Section VIII) is detailed, translating classical CS FLOP counts into comparable metrics and demonstrating up to 60% reduction over EOMP-IDS at high density (N=2048, Q=64). The LLR-CNet design, leveraging BCE loss convergence to exact LLRs, is a clean and falsifiable contribution. The demonstration that the pipeline circumvents error floors from imperfect interferer-count estimation (Fig. 10) is a notable practical result. However, the significance is tempered by the restricted evaluation regime (AWGN channel, perfect equalization), which limits the generalizability of the central 'structural whitener' claim.","major_comments":[{"comment":"§II.C, Eq. (4): The entire evaluation assumes an AWGN channel with H_k=1, which the paper justifies as mirroring 'the effective background noise floor of a post-equalization signal.' This assumes perfect equalization. The central claim that LLR-CNet acts as a 'structural whitener' (§VI) depends on the residual structure being statistically stable and learnable. In a real multipath OFDM system, post-equalization residuals include channel estimation error and residual ISI, whose statistics differ from AWGN. LLR-CNet takes both the cleaned signal and the estimated interference footprint as inputs (Eq. 34); if the signal component carries residual fading structure that LLR-CNet never saw during training, its learned mapping to calibrated soft metrics would be mismatched. The paper does not test this. This is the soft spot: the 'structural whitener' claim is validated only in the regime where","section":null},{"comment":"§VII.D, Eq. (40): The scale-invariance claim assumes the only FFT-size-dependent effect is the 1/√N amplitude scaling. However, the convolutional receptive field (kernel size K=7, padding P=3) operates on a fixed local window. If spectral leakage patterns or the interaction between adjacent interferers change with N (e.g., the density of interferers relative to the grid changes), the learned features may not generalize. The paper demonstrates generalization from N=256 to N=512 (Fig. 10), but this is a modest 2× scaling. A more rigorous test would involve larger scaling factors (e.g., N=1024 or N=2048) to confirm that the local convolutional features remain valid.","section":null},{"comment":"§VII.C, Fig. 9: The claim of >3 dB coding gain under mild interference (SIR=10 dB, Q=12) is attributed to classical algorithms suffering 'signal-peak confusion.' However, the classical baselines (OMP-IDS, EOMP-IDS) are provided with perfect prior knowledge of the interferer count Q in Figs. 9a–c, but it is unclear whether this perfect knowledge is also provided in Figs. 9d–f (SIR=10 dB). If the classical algorithms are disadvantaged by imperfect Q estimation in the mild-interference regime, the comparison may not be fair. The paper should clarify whether the baselines in Fig. 9 operate with perfect or imperfect Q estimation.","section":null}],"minor_comments":[{"comment":"§IV.A, Eq. (20): The notation χ^(0) ∈ R^{N×2} is introduced but the dimension F_l for intermediate feature maps is not clearly defined until later. A brief forward reference would improve readability.","section":null},{"comment":"Table I: The 'Total' parameter count is listed as 14,468, but the sum of the individual layer parameters (480+7,200+2,176+65+2,176+65+2,176+130) equals 14,468. This is correct, but the table would benefit from subtotals for the feature extraction, state conditioning, and output stages.","section":null},{"comment":"§VII.A, Fig. 5: The y-axis label 'Power (dB)' is ambiguous—should specify whether this is relative to the signal power or noise floor.","section":null},{"comment":"§VIII.E, Fig. 11: The intersection point where NBI-CNet becomes more efficient than EOMP-IDS is stated as Q=26, but the figure only shows curves up to Q=64. A marker or annotation at the intersection would aid verification.","section":null},{"comment":"§IX: The conclusion states 'up to a 60% reduction in absolute operations for large grid sizes (N=2048, Q=64),' but Table IV shows the reduction is approximately 59.5% (101.9M vs 41.2M FLOPs). The rounding is acceptable but should be stated precisely.","section":null}],"recommendation":"major_revision","confidential_remarks":"The reader's concern about the AWGN assumption is valid and load-bearing. The paper's framing of LLR-CNet as a 'structural whitener' is strong in the AWGN regime but untested under multipath fading, which is ubiquitous in real OFDM deployments. I recommend the authors add at least one multipath fading evaluation (e.g., EPA or ETU channel) to substantiate the generalizability claim. The novelty disclosure regarding LLR-CNet extending [28] (same first author) is transparently stated and does not appear to be a circularity issue, but the authors should ensure the incremental contribution over [28] is clearly articulated."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this paper proposes a two-stage neural pipeline for narrowband interference mitigation in OFDM. NBI-CNet is a physics-informed CNN that estimates multi-tone interference parameters (gain, fractional frequency offset, phase) in a single forward pass, then reconstructs the interference via Dirichlet kernels and subtracts it. LLR-CNet is a lightweight 1D-CNN that takes both the cleaned signal and the estimated interference footprint and produces calibrated LLRs, trained with BCE loss so the logits converge to true LLRs. The key architectural insight—feeding the estimated interference ˆE alongside the cleaned signal ˜Y to the LLR estimator—is genuinely new and well-motivated. It lets the network localize where residual errors are likely and scale confidence accordingly. The hybrid loss (Eq. 29) with masked parameter losses and a structural sparsity penalty is clean. The complexity analysis is honest: NBI-CNet's advantage over EOMP-IDS only kicks in at Q>26, and they say so plainly. The FFT-size generalization via a 1/√N gain rescaling (Eq. 40) is a nice touch, validated on N=512 after training on N=256. The signal-peak confusion result under mild interference (SIR=10dB, Q=12) where classical greedy algorithms destroy valid payload is a real finding—3dB coding gain there is meaningful. The elimination of error floors from imperfect interferer-count estimation is also a genuine practical advantage. Now the soft spots. The AWGN-only evaluation (H_k=1) is the main limitation, and it's load-bearing for the central claim. LLR-CNet is called a 'structural whitener,' but it's only validated in the regime where the sole non-Gaussian structure comes from NBI residuals. In a real multipath system, post-equalization residuals include channel estimation error and residual ISI with statistics LLR-CNet never saw during training. The paper's justification—that AWGN mirrors the post-equalization noise floor—assumes perfect equalization, which is a strong assumption. This doesn't invalidate the method, but it limits the scope of the whitening claim to the most favorable case. The scale-invariance claim is validated on exactly one unseen FFT size. The train-test mismatch in minimum NBI distance (d_m=2 training vs. d_m=4 evaluation) is unexplained. No code or trained models are released. These are scope-limiting issues, not fatal flaws. The core architecture is sound, the simulations are internally consistent, and the complexity analysis is fair. This paper is for physical-layer and signal processing researchers working on interference mitigation in OFDM, particularly those interested in neural receivers for 5G/6G scenarios. It deserves a serious referee who can push on the multipath question and the generalization claims.","headline":"Joint neural NBI cancellation + soft demodulation for OFDM: solid engineering, but AWGN-only evaluation limits the central whitening claim.","tokens_in":26982,"tokens_out":663,"would_cite":false,"duration_ms":108258,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Neural pipeline cancels narrowband interference without knowing how many interferers exist","keywords":["narrowband interference","OFDM","deep learning","LLR estimation","compressed sensing","Dirichlet kernel","LDPC decoding","interference cancellation"],"falsifier":"A deployment scenario over a realistic frequency-selective fading channel with imperfect channel estimation, where the post-equalization residual structure differs from the AWGN-trained statistics, causing LLR-CNet to produce miscalibrated soft bits and reintroducing the error floors it was designed to eliminate.","tokens_in":26283,"feed_emoji":"📡","tokens_out":953,"duration_ms":220928,"temperature":0.7,"pith_summary":"The paper proposes two neural networks that work together to fix a broken link in OFDM receivers. The first, NBI-CNet, estimates and subtracts narrowband interference in a single pass using a physics-informed convolutional architecture that models spectral leakage via the Dirichlet kernel. The second, LLR-CNet, remaps the non-Gaussian residual left behind after imperfect cancellation into well-calibrated soft bits for the decoder. Together they eliminate error floors that plague classical compressed-sensing pipelines, which suffer from sequential latency, dependence on knowing the exact interferer count, and a mismatch between their structured residuals and the Gaussian assumption of standard demappers.","feed_headline":"Neural pipeline cancels narrowband interference without knowing how many exist","feed_subtitle":"Two-stage deep learning replaces iterative compressed-sensing and Gaussian demapping, eliminating error floors in dense OFDM interference.","key_machinery":"The Dirichlet kernel spectral leakage model, a differentiable analytical reconstruction layer mapping predicted physical parameters to an interference waveform, circular padding in convolutional layers to preserve OFDM spectral wrap-around, binary cross-entropy loss driving network logits to exact LLR convergence, and dynamic conditional routing that evaluates parameter heads only at detected interference locations.","core_discovery":"The central discovery is that the pipeline mismatch between imperfect interference cancellation and Gaussian-soft-demodulation can be resolved by a two-stage neural architecture: a physics-informed estimator that reconstructs the interference waveform analytically from predicted physical parameters (gain, fractional offset, phase), and a lightweight neural LLR estimator trained with binary cross-entropy that naturally converges to exact log-likelihood ratios and acts as a structural whitener for non-Gaussian residuals. The architecture operates without prior knowledge of the active interferer count, generalizes across FFT sizes via a closed-form gain rescaling, and eliminates the error floor","pith_inferences":["If channel estimation errors in real multipath fading channels produce residual structures distinct from those generated by NBI-cancellation errors alone, LLR-CNet's learned residual statistics may not generalize without co-training under realistic channel conditions.","The scale-invariance claim rests on the assumption that the 1/√N amplitude scaling captures all FFT-size-dependent effects; if convolutional receptive fields interact differently with spectral leakage patterns at larger N, a single rescaling factor may be insufficient.","The dynamic conditional routing that bypasses parameter heads at interference-free subcarriers implicitly assumes the gain head achieves near-perfect detection—false negatives would leave interferers unmitigated, a failure mode not characterized in the paper.","Training exclusively under AWGN may limit LLR-CNet's ability to whiten residuals when post-equalization fading artifacts introduce their own non-Gaussian structure, potentially requiring a more diverse training distribution."],"forward_implications":["Receivers in congested spectrum bands (e.g., 6G FR3, IoT, power-line communications) could deploy a single trained model that adapts to varying interference density without retraining or external interferer-count detectors.","The structural-whitener approach to non-Gaussian residuals could generalize to other types of structured interference beyond narrowband tones, such as impulsive noise or multi-user interference, where classical Gaussian demappers also fail.","The scale-invariant design principle—training on a small FFT size and transferring via analytical amplitude rescaling—could reduce the data-generation and training cost for neural physical-layer components across heterogeneous system configurations.","Eliminating the dependency on prior interferer-count estimation removes a single point of failure in classical CS pipelines, making the receiver robust to dynamic environments where interferers appear and disappear on a per-symbol basis."],"fun_headline_variants":["Two-stage neural network cancels OFDM interference and calibrates soft metrics","Physics-informed deep learning eliminates error floors in dense OFDM interference","Joint neural pipeline removes OFDM narrowband interference without active count","Neural interference cancellation hits 0.2 dB of optimal iterative baseline in OFDM","Deep learning resolves pipeline mismatch between NBI cancellation and demodulation"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The evaluation assumes a perfect AWGN channel for the signal of interest, isolating the interference and demodulation modules from channel estimation and equalization errors. Real OFDM systems operate over multipath fading channels, and if residual fading after imperfect equalization interacts with NBI residual structure, the learned residual statistics may no longer match, potentially reintroducing error floors.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage neural network cancels OFDM interference and calibrates soft metrics","Physics-informed deep learning eliminates error floors in dense OFDM interference","Joint neural pipeline removes OFDM narrowband interference without active count","Neural interference cancellation hits 0.2 dB of optimal iterative baseline in OFDM","Deep learning resolves pipeline mismatch between NBI cancellation and demodulation"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":824,"prompt_tokens":729,"completion_tokens":95,"prompt_tokens_details":null},"tokens_in":729,"tokens_out":95,"duration_ms":48644,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T02:24:27.826763+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"A deployment scenario over a realistic frequency-selective fading channel with imperfect channel estimation, where the post-equalization residual structure differs from the AWGN-trained statistics, causing LLR-CNet to produce miscalibrated soft bits and reintroducing the error floors it was designed to eliminate.","supporting_citations":[],"review_version":1}