{"id":"d67a4cc7-4c4d-4c49-8687-c449f9d461ba","arxiv_id":"2608.06230","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A hybrid scattering-transform CNN conditioned on modulation and coding scheme predicts 5G NR PUSCH CRC outcomes from received I/Q samples before LDPC decoding, reaching 95% accuracy and 3.91% missed-detection rate in simulation.","lead":"This paper trains a small neural network to predict whether a 5G data packet will pass its error check before the receiver finishes decoding it, using only the raw received signal and channel estimates. In simulations of 5G New Radio uplinks, the best version reaches 95% accuracy and rarely misses a failed packet, which could let base stations retransmit nearly a full millisecond earlier.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The RTT-reduction claim relies on unmeasured narrowband mini-slot performance; the evaluated full-bandwidth, 13-symbol configuration does not match the target URLLC scenario, and measured inference time exceeds the slot-boundary budget.","rationale":"The reader's verdict of CONDITIONAL is appropriate. I identify a different weakest link than the reader: the transfer from the evaluated full-bandwidth, 13-symbol PUSCH configuration to the target URLLC narrowband mini-slot scenario. The paper's own Section 4.5 admits that the narrowband latency is 'expected' but not measured, and that admission is a limitation the manuscript should have flagged more strongly. The reader's chosen weakest assumption about the single-DMRS LS channel estimate is plausible but less decisive, because the received data symbols y are also available at full resolution and carry channel magnitude and SNR information; the network is not solely dependent on the interpolated LS estimate. The central accuracy results (95% accuracy, 3.91% MDR, 10.00% FAR) are internally consistent and the ablations (scattering vs. pure CNN, MCS conditioning) are informative. However, those results were obtained on 273-PRB, 13-symbol slots, while the motivating deployment scenario is a 7-symbol mini-slot with 25–50 PRBs. The measured full-bandwidth inference time (0.299 ms real compute) already exceeds the available 0.142 ms slot-boundary window, so the deployment claim rests entirely on an unverified scaling assumption. A concrete narrowband/mini-slot experiment would either confirm the RTT reduction or show that the model needs retraining and possibly different front-end parameters. Until that is done, the paper's headline operational benefit should be treated as conditional, matching the reader's verdict.","tokens_in":13122,"tokens_out":12006,"duration_ms":119169,"concrete_test":"Generate 25-PRB, 7-symbol PUSCH slots with front-loaded DMRS using the same Aerial pipeline, and run the FFT+CNN+MCS TensorRT engine on the L4 GPU. Measure end-to-end inference time and classification MDR/FAR on a held-out out-of-distribution test set, with decision thresholds tuned on a separate validation set. If the inference time exceeds the remaining-symbol budget (about 142 us at 30 kHz SCS) or if accuracy/MDR/FAR degrade materially relative to the full-bandwidth results, the central RTT-reduction claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's operational claim—reducing HARQ RTT from 5 ms to 0.5 ms—depends on the predictor completing within the ~142 us left after a 7-symbol mini-slot and maintaining accuracy on a narrowband, shorter resource grid. However, all experiments use a 273-PRB, 13-symbol PUSCH slot (Table 2). The measured TensorRT real compute for FFT+CNN is 0.299 ms (Section 4.5), which exceeds the 0.142 ms slot-boundary budget; the paper's response is an extrapolation ('expected to be proportionally shorter', Section 4.5) with no narrowband latency measurement and no mini-slot accuracy evaluation. Because the model is trained and evaluated on 13-symbol grids, its behavior on 7-symbol mini-slots is untested. This concern is more load-bearing than the single-DMRS LS-interpolation issue: even if the LS expansion loses fine spectral structure, the received data symbols themselves (which include the channel response multiplied by known-constellation data) carry substantial channel-state information, providing a redundant path for the network to infer decodability. Without a narrowband, mini-slot measurement, the central latency and deployment claims are unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a neural CRC predictor for 5G NR PUSCH that takes the received I/Q resource grid and LS channel estimates, extracts first-order scattering features via either Gabor or FFT-based fixed front-ends, and classifies CRC pass/fail with a compact CNN, optionally conditioned on the MCS index. On an NVIDIA Aerial simulation with 30,000 training and 20,000 test slots, the scattering front-ends are reported to substantially outperform a pure CNN, MCS conditioning reduces the false-alarm rate of the scattering models, and the best operating point is FFT+CNN+MCS with 95% accuracy, 3.91% missed-detection rate, and 10.00% false-alarm rate. An evidential extension with a conservative uncertainty rule is shown to reduce MDR to 0.76% at 42.24% FAR. The paper claims this enables reducing effective HARQ round-trip time from 5 ms to 0.5 ms for URLLC mini-slots.","tokens_in":13385,"tokens_out":6203,"duration_ms":62762,"significance":"If the operational claims held, the paper would demonstrate a practically useful mechanism for early retransmission scheduling in URLLC. The empirical core is internally consistent: the comparison of scattering front-ends against a pure CNN uses large margins, the fixed front-ends add essentially no trainable parameters, the TensorRT compatibility is a concrete strength, and the evidential-uncertainty extension is a reasonable single-pass alternative to sampling-based methods. The main weakness is that the deployment-level claim is not supported by the measurements: all experiments use a full-bandwidth, 13-symbol slot, whereas the motivating scenario is a 7-symbol mini-slot with a 142 us processing budget, and the measured FFT+CNN real compute time of 0.299 ms already exceeds that budget. The lack of multi-seed confidence intervals also weakens the finer-grained comparisons between front-ends and the evidential threshold results. These issues are fixable with additional experiments or a qualified reframing.","major_comments":[{"comment":"The central deployment claim—that the predictor reduces HARQ RTT from 5 ms to 0.5 ms by completing within the ~142 us left after a 7-symbol mini-slot—is not supported by any presented measurement. All experiments use a 273-PRB, 13-symbol PUSCH slot (Table 2), and the measured TensorRT real compute for FFT+CNN is 0.299 ms (Section 4.5), which exceeds the 0.142 ms slot-boundary budget. The statement that inference time is 'expected to be proportionally shorter' for narrowband URLLC allocations is an extrapolation; no narrowband latency measurement and no mini-slot accuracy evaluation are reported. Because the model is trained and evaluated on 13-symbol grids, its behavior on shorter or narrower grids is untested. Please either measure the actual mini-slot configuration or restrict the abstract's and conclusion's RTT-reduction claims to configurations consistent with the experiments.","section":"Section 1.1 and Section 4.5"},{"comment":"All headline numbers are single point estimates from one train/test split, with no seeds, confidence intervals, or significance tests. The front-end comparisons used to justify design choices—for example, FFT+CNN versus Gabor+CNN with MCS conditioning (MDR 3.91% versus 2.90%; FAR 10.00% versus 18.28%)—include differences of a few percent that could be within run-to-run variability. In addition, the evidential uncertainty threshold tau_u = 0.025 in Table 3 appears to be chosen on the test set; if so, the reported 0.76% MDR is an optimistic in-sample selection result. Please report means and standard deviations over multiple seeds and use a validation-based procedure for threshold selection.","section":"Section 4.2 and Table 3"},{"comment":"The conservative decision rule of Eq. (29) effectively reclassifies all high-epistemic-uncertainty samples as failures; at tau_u = 0.025, 19.3% of test samples are flagged and FAR rises from 10.00% to 42.24%. The paper justifies this as acceptable by comparing expected retransmission count with blind repetition, but this ignores that each false alarm consumes a scheduling opportunity and adds latency or jitter for a TB that would have succeeded. A proper evaluation should report the MDR-versus-FAR operating curve over tau_u and compare at matched FAR, or simulate the resulting HARQ process, before claiming that the evidential rule provides the best safety margin for URLLC.","section":"Section 4.6"}],"minor_comments":[{"comment":"The text says 'The modulus in (9)' but the modulus operation is defined in Eq. (11); Eq. (9) is the concatenation of received symbols and channel estimates.","section":"Section 3.2"},{"comment":"The bar chart labels overlap the bars, making the figure difficult to read; the normalized-latency series duplicates information from Section 4.5 and should be shown in a separate panel.","section":"Fig. 3"},{"comment":"The test set is described as out-of-distribution in delay spread and UE speed, but Table 2 lists the same delay-spread range for training and inference; please specify exactly which delay-spread/UE-speed combinations are held out and whether SNR or channel model realizations differ.","section":"Table 2 and Section 4.1"},{"comment":"Several references have formatting errors (e.g., [2] lists 'Huttunen, J. M. J. , M.'), and the manuscript does not include a data/code availability statement, which would aid reproducibility of the numerical claims.","section":"References"},{"comment":"The statement that a false alarm 'wastes resources, but far less than blind repetition' assumes that one false alarm costs exactly one retransmission; under configured-grant or retransmission-limit operation the cost may be higher, so the claim should be qualified.","section":"Section 2.4"}],"recommendation":"major_revision","confidential_remarks":"The empirical core is sound as a simulation study, but the paper's most prominent operational claim—RTT reduction from 5 ms to 0.5 ms—is not supported by the measurements and needs either new mini-slot/narrowband experiments or careful qualification. The single-run evaluations and test-set-selected uncertainty threshold are additional reproducibility concerns. If the journal values deployment-validated claims, the current version is not ready; if it values the simulation result alone, the framing should be adjusted accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hey,\n\nQuick take: this is a genuinely new application—predicting post-LDPC CRC outcome directly from the received I/Q grid and LS channel estimates, before decoding. The authors know the neural-receiver literature and borrow its input conventions cleanly. The scattering front-end plus MCS conditioning is a sensible combination, and the simulation results are internally consistent: both scattering models beat the plain CNN by a wide margin on MDR, and MCS conditioning roughly halves FAR for the FFT variant. The framing of MDR vs FAR in terms of URLLC retransmission cost is correct and useful. This is the strongest part of the paper.\n\nThe soft spots are real and mostly about evidence quality. There are no error bars or multiple seeds; the decision threshold and the epistemic threshold are tuned on the test set; and no code or data are released. I can't falsify the central claim, but I can't fully trust the specific numbers either.\n\nThe bigger problem is the latency/deployment claim. The paper's motivating scenario is a 7-symbol URLLC mini-slot, and the stated budget is ~142 us before the slot boundary. But the model is trained and evaluated only on a 13-symbol, 273-PRB PUSCH slot. Measured TensorRT real compute for FFT+CNN is 0.299 ms—more than twice the budget. The paper hand-waves that narrowband inference will be proportionally shorter, but that is not a measurement, and there is no evaluation on 7-symbol grids at all. So the 5ms-to-0.5ms HARQ RTT reduction is unsupported by the presented data. This is a load-bearing flaw for the deployment claim, not a nitpick.\n\nOne mitigating note: I think the stress-test's worry about single-DMRS LS interpolation is mostly a non-issue. The received data symbols themselves carry channel-state information, so the network has a redundant path to infer decodability. Not a fatal concern.\n\nBottom line: the core idea deserves a serious referee. It is a legitimate research direction with a coherent simulation study. What it needs is a revision that reports variance, releases the artifacts, and produces actual narrowband mini-slot latency and accuracy measurements. As it stands, I'd send it to review with major revision requested, not desk reject.","headline":"Early CRC prediction for URLLC is a genuinely new and worthwhile idea, but the latency claim is unsupported by the full-bandwidth measurements.","tokens_in":13909,"tokens_out":2069,"would_cite":true,"duration_ms":19206,"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 neural network can predict whether a 5G uplink block will pass its CRC check before decoding finishes, cutting retransmission delay from 5 ms to 0.5 ms.","keywords":["5G NR","PUSCH","URLLC","CRC prediction","scattering transform","evidential deep learning","HARQ latency","MCS conditioning"],"falsifier":"Feed the FFT+CNN+MCS predictor test slots in which a deep narrowband fade falls between DMRS subcarriers, and compare its missed-detection rate with that of the same model retrained on a two-DMRS pilot pattern; a sharp degradation in the single-DMRS version would show that the interpolation assumption, not the classifier, is the limiting component.","tokens_in":12871,"feed_emoji":"📡","tokens_out":6706,"duration_ms":56442,"temperature":0.7,"pith_summary":"This paper proposes that a neural network can predict the cyclic redundancy check (CRC) outcome of a 5G NR PUSCH transport block directly from the received I/Q resource grid and least-squares channel estimates, bypassing equalization, demodulation, and LDPC decoding. If the claim holds, a base station would know before the slot boundary which blocks fail, allowing it to schedule a retransmission in the very next uplink slot and reduce the effective HARQ round-trip time from 5 ms to 0.5 ms. The design combines a fixed scattering front-end with a compact CNN and an auxiliary MCS-index input; the best configuration reports 95% accuracy, 3.91% missed-detection rate, and 10.00% false-alarm rate on an out-of-distribution test set. An evidential extension using a conservative uncertainty rule further lowers the missed-detection rate to 0.76%.","feed_headline":"Neural predictor cuts 5G retransmission delay from 5 ms to 0.5 ms","feed_subtitle":"A lightweight CNN reads raw uplink symbols and channel estimates to flag failed transport blocks before LDPC decoding ends.","key_machinery":"The load-bearing mechanism is a fixed first-order scattering front-end that converts the received signal and channel estimate into phase-invariant, multi-scale energy features. The Gabor realization applies J=2 scales and L=4 orientations as spatial depthwise convolutions; the FFT realization multiplies the input's two-dimensional Fourier transform by constant-quality-factor bandpass masks (Q=2) with center frequencies geometrically spaced from f_min=0.02 to f_max=0.4 and recovers the features by inverse transform and modulus. Both versions feed a compact CNN classifier with roughly 7,400 trainable parameters, and the MCS index enters as a learned embedding concatenated with the pooled features before the final sigmoid. The evidential extension replaces the sigmoid with Dirichlet concentration parameters, enabling a single-pass decomposition of total uncertainty into aleatoric and epistemic parts and a conservative decision rule for predicting CRC failure.","core_discovery":"The central discovery is that the post-decoding CRC verdict of a 5G NR PUSCH transport block is learnable from raw received symbols plus a least-squares channel estimate, without running the conventional equalization–demodulation–decoding chain. The paper shows that a fixed first-order scattering front-end, realized either with Gabor filters or with FFT-domain bandpass masks at geometrically spaced scales, supplies the multi-scale time-frequency energy features a compact CNN needs: replacing a pure CNN with either front-end raises accuracy from 82% to 90% and cuts the missed-detection rate from 19.20% to below 4.5%. Conditioning on the modulation and coding scheme (MCS) index lets the model adapt its decision boundary to the operating code rate, cutting the FFT variant's false-alarm rate from 26.21% to 10.00% while raising accuracy to 95%. The evidential output head, using a conservative rule that predicts failure when epistemic uncertainty exceeds a threshold, reaches 0.76% missed detections, and both front-ends execute within the slot processing budget on an edge-class GPU.","pith_inferences":["A testable extension is to compare the single-DMRS expanded channel estimate with denser pilot configurations; if accuracy improves substantially, the interpolation step, not the network, is the current ceiling.","The same early-CRC logic should transfer to downlink reception or other CRC-protected channels, where the same timing pressure exists.","Because false alarms waste resources but missed detections are silent, operators could adapt the thresholds per quality-of-service flow rather than per network.","The 5 ms-to-0.5 ms latency claim assumes the predictor's own runtime fits between the mini-slot's end and the slot boundary; measuring end-to-end MAC-scheduler integration would confirm whether the theoretical gain survives protocol overhead."],"forward_implications":["A gNB can schedule a retransmission in the next uplink slot, cutting effective HARQ round-trip time from 5 ms to 0.5 ms for mini-slot URLLC.","Configured-grant blind repetitions (2–8×) can be replaced by triggered retransmissions, saving air-interface resources while preserving deterministic latency.","The same trained evidential predictor can serve both eMBB (standard rule) and URLLC (conservative rule) by switching decision thresholds at inference.","MCS conditioning makes one predictor span six code rates; the FFT front-end's real compute is 0.299 ms on a full 273-PRB grid and scales down with URLLC-sized allocations.","The decision thresholds (tau and tau_u) give operators deployment-time knobs to trade missed detections against unnecessary retransmissions."],"supporting_citations":[{"why":"Establishes the protocol and processing-delay bottleneck, specifically scheduling and feedback timing, that early CRC prediction targets.","marker":"[1]"},{"why":"Supplies the neural-receiver input convention (received I/Q plus LS channel estimates) adopted by the predictor.","marker":"[2]"},{"why":"Provides the invariant scattering convolution framework from which the first-order energy features are drawn.","marker":"[8]"},{"why":"Gives the FFT-domain fast scattering computation that the FFT front-end realizes in the frequency domain.","marker":"[9]"},{"why":"Provides the evidential deep learning loss and uncertainty decomposition used in the extension.","marker":"[10]"},{"why":"The GPU-accelerated RAN simulation framework that generates the training and test data, CRC ground truth, and TensorRT latency numbers.","marker":"[11]"}],"fun_headline_variants":["Neural CRC predictor skips decoding to slash 5G URLLC latency","CNN reads raw symbols to flag 5G block errors early","95% accurate neural CRC prediction for 5G PUSCH","Early link adaptation via neural CRC prediction in 5G"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The network must learn decodability from a single-DMRS least-squares channel estimate expanded across the full resource grid; if that expansion hides the fine-grained per-subcarrier SNR or frequency-selective fading structure that determines LDPC decodability, the predictor cannot recover it.","fun_headline_variants_meta":{"raw":{"variants":["Neural CRC predictor skips decoding to slash 5G URLLC latency","CNN reads raw symbols to flag 5G block errors early","95% accurate neural CRC prediction for 5G PUSCH","Early link adaptation via neural CRC prediction in 5G"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1431,"prompt_tokens":1044,"completion_tokens":387,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":312}},"tokens_in":660,"tokens_out":387,"duration_ms":4162,"temperature":1.0,"reasoning_tokens":312,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:58:32.359577+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed the FFT+CNN+MCS predictor test slots in which a deep narrowband fade falls between DMRS subcarriers, and compare its missed-detection rate with that of the same model retrained on a two-DMRS pilot pattern; a sharp degradation in the single-DMRS version would show that the interpolation assumption, not the classifier, is the limiting component.","supporting_citations":[{"cited_title":"M., and Hassanieh, H.: Ultra-Reliable Low-Latency in 5G: A Close Reality or a Distant Goal?","cited_arxiv_id":null,"evidence_quote":"Establishes the protocol and processing-delay bottleneck, specifically scheduling and feedback timing, that early CRC prediction targets."},{"cited_title":"In: IEEE Transac- tions on Pattern Analysis and Machine Intelligence, vol","cited_arxiv_id":null,"evidence_quote":"Provides the invariant scattering convolution framework from which the first-order energy features are drawn."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The GPU-accelerated RAN simulation framework that generates the training and test data, CRC ground truth, and TensorRT latency numbers."}],"review_version":2}