{"id":"36baf5b6-d4fb-41ea-8b26-d8f21b014be0","arxiv_id":"1908.00460","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Adding a residual-learning denoiser before a neural decoder improves BER for a short polar code over a plain neural decoder at comparable latency.","lead":"Researchers built a neural network decoder for polar codes that first denoises the received signal with a residual-learning module, then decodes it. The method improves bit-error rate over a plain neural decoder at similar latency, and the best version approaches the classic SC decoder while running much faster.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The latency comparison in Section V-D is unauditable: Fig. 9 omits platform, SC implementation, framework, batch size, and run protocol, so the 'comparable latency' and over-100x speedup claims are not established.","rationale":"I read the paper as trying to establish two intertwined claims: (1) a residual-learning denoiser plus jointly trained neural decoder improves BER over a plain NND of similar parameter count, and (2) this is achieved at comparable or better latency, with MLP-RNND being a practical low-latency alternative to SC. The BER claim is supported by the curves in Fig. 8 and the SNR/PDF evidence in Figs. 6-7, so I do not see a fatal flaw in the architecture or training procedure itself. The weakest link is the latency evidence. Figure 9 is the sole support for the 'comparable latency' and 'over 100 times faster than SC' statements, yet the measurement methodology is entirely absent: no platform, no SC implementation details, no batch size, no averaging, and no code. This is a load-bearing concern because the abstract, introduction, and conclusion all use the low-latency motivation to position the contribution; if the timing comparison is an artifact of implementation or hardware, the practical relevance claim is unverified. The reader's weakest_assumption identified the same issue, so I agree. A conditional verdict with a request for a reproducible benchmark is appropriate; the BER claim should not be rejected on this basis, but it should also not be accepted as a complete low-latency solution without the missing measurement details.","tokens_in":7764,"tokens_out":11592,"duration_ms":132013,"concrete_test":"Run a controlled benchmark on a single machine: implement SC in optimized C/C++ with the same compiler flags, run MLP-NND and MLP-RNND in PyTorch or TensorFlow on the same CPU/GPU, and report per-codeword wall-clock time in both single-shot and fixed-batch modes over at least 1e4 trials, along with hardware and software versions. If MLP-RNND is not within a small factor of MLP-NND latency, or not over 100 times faster than optimized SC, the latency portion of the central claim fails; the BER comparison should then be evaluated independently of the speedup claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed RNND achieves better BER than the corresponding NND under comparable latency, and that MLP-RNND runs over 100 times faster than SC. The BER half of this claim is supported by the simulation curves in Fig. 8, but the latency half rests entirely on the computation-time comparison in Section V-D and Fig. 9. The paper gives no details about the hardware (CPU/GPU model), software framework, SC implementation language or optimization level, batch size, number of trials, or whether neural-network inference was batched. Without these details, the reported 'more than 100 times faster' result could reflect a naive or unoptimized SC reference running on a CPU against batched GPU inference for the neural decoders, rather than an algorithmic advantage. Similarly, the statement that RNNDs run 'slightly slower' than NNDs is not quantified. Since the abstract and conclusion explicitly motivate the method by low latency, this missing measurement protocol is the least secure condition for the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a residual neural network decoder (RNND) for short polar codes. The RNND inserts a residual-learning denoising module before a neural network decoder and trains both jointly with a multi-task loss L = L_denoise + L_decode (Eq. 6). Using (16,8) polar codes with BPSK over AWGN, the authors compare MLP-, CNN-, and RNN-based RNNDs against their NND counterparts and against successive cancellation (SC). Numerical results show that each RNND improves BER relative to its NND counterpart, with MLP-RNND approaching SC performance, and the paper claims that MLP-RNND runs more than 100 times faster than SC.","tokens_in":7977,"tokens_out":5356,"duration_ms":58472,"significance":"If the reported results hold, this is a modest but useful empirical contribution to neural decoding of short polar codes. The architecture is simple, the comparison across three neural-network families is informative, and the parameter-matched NND baselines are a fair design choice. The paper does not contain mathematical derivations or machine-checked proofs, and the performance claims are based on simulations only; nevertheless, the proposed loss and architecture are concrete and falsifiable, and the reported BER gains are consistent across all three architecture families. The main weakness is that the paper's key motivation, low latency, rests on a computation-time comparison that currently lacks sufficient measurement detail to be verified.","major_comments":[{"comment":"The computation-time comparison that underlies the abstract's low-latency claim is not auditable. The manuscript does not report the CPU/GPU model, the software framework, the SC implementation details (language, optimizations, whether a standard or optimized version was used), the batch size, the number of repeated trials, or whether neural-network inference was batched while SC was run serially. As written, the claimed 'more than 100 times faster' figure and the statement that RNNDs run only 'slightly slower' than NNDs cannot be reproduced or verified, and the comparison may not be apples-to-apples. Please document the measurement protocol and provide absolute timings for all decoders on the same platform.","section":"V-D, Fig. 9"},{"comment":"The BER curves in Fig. 8 appear to be single simulation traces without error bars, confidence intervals, or multiple training seeds. Since the headline gain of MLP-RNND over MLP-NND is about 0.2 dB at BER 10^-4 and the curves are close in some SNR regions, the authors should report the number of test blocks and either error bars or mean plus/minus standard deviation over several independent training runs, so that the reader can assess whether the reported gains are statistically meaningful.","section":"V-C, Fig. 8"}],"minor_comments":[{"comment":"In the sentence 'Since Ldenoise andLdenoise are both continuous and differentiable everywhere', the second loss should be Ldecode; the current text contains a typographical error.","section":"IV-C"},{"comment":"In the training procedure, 'Leaning rate' should be 'Learning rate'.","section":"IV-D"},{"comment":"The sentence describing Fig. 6 contains 'test-SNR = dB', which appears to be missing a value; please specify the SNR value and describe how the SNR after denoising is computed.","section":"V-B"},{"comment":"The statement that 'RNNDs run slightly slower than NNDs' is not quantified; please give the measured inference times so that the reader can judge the magnitude of the difference.","section":"V-D"},{"comment":"The concluding claim that the RNND can be extended to longer code lengths is speculative, because no experiments or training-complexity analysis for longer polar codes are provided; please temper or support this statement.","section":"VI"},{"comment":"There is a typo in 'trainsmitted'; it should be 'transmitted'.","section":"III"}],"recommendation":"major_revision","confidential_remarks":"The empirical contribution is modest and the paper is written in a short-form style. The main obstacle to publication is the latency benchmark: the paper's central motivation is low latency, but the reported speedup lacks sufficient measurement detail to be checked. If the authors can supply a complete measurement protocol, error bars for the BER results, and absolute timings for SC and all neural decoders on the same platform, I would view this as a publishable short paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the actual contribution is a modest but real one—putting a residual-learning denoiser in front of a neural polar decoder and training the two jointly with a sum of denoising and decoding losses. For the tested case (N=16, R=1/2, AWGN), the BER curves in Fig. 8 show a few tenths of a dB gain over a plain NND at BER 1e-4, and the comparison against equal-parameter NNDs is the right way to isolate the effect. That part is credible.\n\nWhere the paper falls down is the latency claim. The abstract and conclusion sell the method on low latency, but the entire '100x faster than SC' and 'comparable latency' argument rests on Fig. 9, which has zero measurement detail: no hardware model, no framework, no SC implementation notes, no batching information, no trial count. That is a load-bearing missing protocol, not a minor omission. If the SC reference is a naive CPU loop and the neural nets are batched on a GPU, the speedup claims are meaningless. This is the same point the stress-test note makes, and reading the paper confirms it.\n\nOther soft spots: no error bars or multiple-seed statistics, no baseline with a non-residual denoiser (so you cannot tell how much of the gain comes from residual learning vs. just adding a trainable front-end), and the scope is genuinely narrow—short polar codes only, with no evidence the approach scales. The paper also has a small typo in Section V-B where a test-SNR value is missing, and the PDF figure is more illustrative than quantitative. These are minor next to the latency issue.\n\nThe citation pattern looks fine; the authors build on the NND literature and the denoising literature without overclaiming novelty. The central BER result is plausible and the multi-task loss is clean.\n\nWho is this for: someone working on neural decoding of short codes who wants a quick, cheap way to squeeze out a bit of performance. It is not going to change the landscape for practical polar decoding. It deserves a serious referee only if the latency comparison can be made reproducible—otherwise the paper's main motivation is unverifiable. I would accept it for peer review with major revisions, specifically asking for the measurement protocol, repeated trials, and an ablation with a plain (non-residual) denoiser. That is enough to turn a plausible simulation study into a solid one.","headline":"A genuinely new but small combination—residual denoising plus neural polar decoding—with a real BER gain at N=16, undermined by an unauditable latency comparison that the paper's own motivation rests on.","tokens_in":8519,"tokens_out":1230,"would_cite":false,"duration_ms":15175,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-14T15:54:45.450959+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}