{"id":"0a6f0466-c69a-491a-bb37-c378c1db619c","arxiv_id":"2506.10166","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"An attention-based neural polar decoder with a block-level loss improves simulated BER and BLER for a (256,37) polar code, and a CRC-guided ensemble variant reports further gains.","lead":"DeepPolar+ adds self-attention and a joint bit/block loss to a neural polar decoder, reporting small BER and BLER gains over successive cancellation for a length-256 code. A CRC-guided multi-decoder variant reports larger gains, but the comparison does not account for the information rate lost to the CRC bits.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SMART gains are confounded by CRC payload reduction: Eq. (12) puts r CRC bits inside the 37 information positions, so DP+SMART carries only 37−r user bits; results are plotted vs SNR rather than Eb/N0, making the headline 0.5/0.6 dB gains potentially rate-driven.","rationale":"The reader's weakest assumption exactly identifies the rate confound from Eq. (12): CRC bits appended before encoding reduce the user payload unless the code dimension is increased, but the paper keeps calling it a (256,37) code. This is the most load-bearing issue because the headline SMART gains are the paper's main quantitative contribution. The comparison is at channel SNR, not Eb/N0, so a lower-rate code has an inherent energy-per-bit advantage; without an equal-payload baseline such as CRC-aided SCL, the reported 0.5/0.6 dB gains cannot be attributed to the architecture. The non-SMART DeepPolar+ claim of breaking the BER-BLER trade-off is also fragile: the 0.08 dB BLER margin over SC is well within typical Monte Carlo noise and no error bars or repeated-seed results are provided. While the architecture and training-loss ideas may have merit, the central claim as stated is not supported by the evidence. I agree with the reader's reject verdict and find no reason to adjust it.","tokens_in":8640,"tokens_out":6694,"duration_ms":81939,"concrete_test":"Re-run the SMART comparison with equal information rate: set user payload to 37 bits for all schemes, e.g., use a (256,45) DeepPolar code for 8-bit CRC (37 info + 8 CRC) or replace SC/DeepPolar baselines with CRC-aided SCL using the same 8-bit CRC and 29-bit payload, and plot BER/BLER vs Eb/N0 instead of SNR. If the SMART gain at 10^-4 BER / 10^-3 BLER drops below 0.1 dB, the headline claim is a rate artifact. Also report confidence intervals from at least 5 independent training runs.","verdict_should_be":"REJECT","load_bearing_attack":"The central quantitative claim is the 0.5 dB BER / 0.6 dB BLER gain of DP+SMART over DeepPolar for a (256,37) code. Eq. (12) states x = g+_phi([u, CRC(u)]) ∈ R^n with u ∈ {0,1}^k. Since the encoder's information block for the (256,37) code has k=37 positions, appending r CRC bits means r of those positions are not user payload. With the 8-bit CRC used for the reported gains, the true message length is 29 bits (R=29/256), not 37 bits (R=37/256); with 3-bit CRC it is 34 bits. The paper never states this effective rate and plots BER/BLER against channel SNR (1/σ²) rather than Eb/N0, so the SNR-axis gains include the energy-per-information-bit advantage of a 21.6% lower rate. No CRC-aided SCL or other equal-payload baseline is provided. Thus the SMART improvements cannot be attributed to the neural decoder; they may be a rate artifact. The non-SMART DeepPolar+ improvement of 0.08 dB in BLER also has no confidence intervals, so even the trade-off-breaking claim rests on a possibly noise-level margin.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DeepPolar+, an attention-augmented neural successive cancellation decoder for DeepPolar codes, together with a loss that adds a soft block-error term to the bit-wise cross-entropy and a multi-model SNR-matched decoding framework with CRC verification (DP+SMART). For a (256,37) code on the AWGN channel, it reports BER and BLER improvements over successive cancellation and DeepPolar, including up to 0.5 dB BER gain at 10^-4 and 0.6 dB BLER gain at 10^-3 for DP+SMART, and claims that the architecture removes the BER-BLER trade-off while reducing training epochs by 75%.","tokens_in":8958,"tokens_out":6215,"duration_ms":66282,"significance":"If the reported gains survive equal-rate and equal-complexity comparisons, DeepPolar+ would be a meaningful advance in neural channel coding: it provides a public implementation, a detailed architecture description, and several useful analyses (training convergence, codeword-distance statistics, and SNR specialization). The paper is clearly written and the experiments are easy to reproduce in principle. However, the central quantitative claims are currently confounded by an unstated payload-rate reduction in the SMART scheme, by the absence of CRC-aided list-decoding baselines, and by test-set model selection. The contribution is therefore not yet established at the level claimed.","major_comments":[{"comment":"The CRC augmentation is rate-confounding. Equation (12) writes x = g+_phi([u, CRC(u)]) with u in {0,1}^k, but the encoder g+_phi maps {0,1}^k to R^n; if the concatenated vector has k+r entries, the encoder domain is inconsistent. If k is instead the total number of encoder input bits including the CRC, then the true payload is k-r: 29 bits for the 8-bit CRC and 34 bits for the 3-bit CRC, not 37 bits. All curves in Fig. 2a are plotted against SNR (1/sigma^2), not Eb/N0, so the reported 0.5 dB BER and 0.6 dB BLER gains include the energy-per-information-bit advantage of a 21.6% lower rate. The effective information rate of every SMART variant must be stated, and equal-payload comparisons (e.g., Eb/N0 plots or (256,29) and (256,34) reference codes) are required before any gain can be attributed to the neural decoder.","section":"III-C, Eq. (12)"},{"comment":"No CRC-aided successive cancellation list (CA-SCL) baseline is provided. Since DP+SMART uses CRC verification and an ensemble of decoders, the relevant comparison is CA-SCL with the same CRC polynomial and the same effective rate. Without this baseline, the reported gains over plain SC decoding could be explained by the CRC's error-detection capability and list diversity rather than by the attention-enhanced neural decoder. The paper's conclusion that DeepPolar+ is superior to existing polar decoding approaches is unsupported until such a baseline is included.","section":"IV-E, Fig. 2a"},{"comment":"The SMART SNR-pair set appears to be selected after inspecting Fig. 2d, which reports performance on the same SNR range used for evaluation. This constitutes model selection on the test set and inflates the reported DP+SMART gains. The authors need to specify how the five SNR pairs were chosen (e.g., on a separate validation set) and, ideally, average over multiple random choices of candidate pairs to assess selection bias.","section":"IV-E, Fig. 2d"},{"comment":"The key trade-off-breaking claim rests on a 0.08 dB BLER improvement at 10^-3 over SC decoding, yet no error bars, confidence intervals, or multiple-seed results are reported for any curve. Given typical Monte Carlo uncertainty at these error rates, this margin may be noise. The paper should report the number of simulated blocks, standard errors, or confidence bands for at least the headline BER and BLER values.","section":"IV-B"}],"minor_comments":[{"comment":"The naming is inconsistent: 'DeepPolar', 'DeepPolar+', and 'Deeppolar' are used interchangeably in Fig. 2a and the text; please unify notation.","section":"Throughout"},{"comment":"Equation (12) has a typo: 'u ∈ 0,1^k' should be 'u ∈ {0,1}^k', and the CRC polynomial should be defined with its standard generator before use.","section":"III-C"},{"comment":"The empirical choice of kernel size ℓ=16 as √N is not justified; please provide an ablation or a reference for this heuristic.","section":"IV-A"},{"comment":"The Gaussian-distance analysis would be stronger with a quantitative divergence metric (e.g., KL divergence) and a comparison to the original DeepPolar distribution; the current visual similarity claim is qualitative.","section":"IV-C"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the public code and clear writing are assets, but the rate confound in the SMART evaluation is serious. If the equal-rate re-analysis confirms the gains, the paper may be publishable after adding CA-SCL baselines and error bars; if not, the contribution reduces to an incremental architecture. I would also ask the editor to check the relationship to the authors' prior DKO work and the DeepPolar baseline to ensure novelty is clear."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know that the headline SMART gains are not credible as stated. The CRC handling in Eq. (12) changes the effective information rate, and the paper never reports the true payload length. For the (256,37) code, an 8-bit CRC means only 29 user bits per codeword, so the 0.5/0.6 dB gains over the (256,37) SC baseline are at least partly rate-driven. The curves are plotted against SNR, not Eb/N0, which hides that advantage. This is the load-bearing flaw.\n\nThat said, the paper is not empty. The attention-enhanced decoder is a sensible extension of DeepPolar, and the joint BCE + soft block loss is a reasonable way to nudge BLER. The reported 75% reduction in training epochs is a useful practical result, and the codeword distance analysis is a nice sanity check. The non-SMART DeepPolar+ results are modest but plausible: 0.4 dB over SC at BER 1e-4 and matching SC's BLER within 0.08 dB. If the authors had stopped there, the paper would be a modest but honest incremental contribution.\n\nThe soft spots beyond the CRC confound: there are no error bars, so the 0.08 dB BLER 'improvement' over SC could be noise. There is no CRC-aided SCL baseline, which is the standard practical polar decoder and would likely close much of the gap. The SMART model set was chosen after looking at Fig. 2d, which reports performance on the same evaluation SNR range, so the ensemble comparison is tuned. The paper acknowledges this indirectly, but it still undermines the claim of a general technique.\n\nWho is this for? Researchers working on neural polar codes and learned decoders. They will find the architecture and training loss worth a look, but the evaluation needs to be redone at equal rate and against SCL. I would not cite it in its current form.\n\nMy recommendation: desk reject is too harsh, but accept for peer review with a strong request to fix the rate comparison, add error bars, and include an SCL baseline. If those are addressed, the attention and loss contributions could become a citable result. As is, the central 'breaking the BER-BLER trade-off' claim is not supported.","headline":"SMART gains are likely rate-driven from hidden CRC overhead; the attention decoder and joint loss are plausible but the paper's central claim needs equal-rate comparison.","tokens_in":9481,"tokens_out":2488,"would_cite":false,"duration_ms":27101,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94B35","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"DeepPolar+ sets out to prove that a neural polar code can improve block error rate and bit error rate simultaneously, ending the BER-BLER trade-off that limited DeepPolar.","keywords":["polar codes","neural channel coding","self-attention decoder","BER-BLER trade-off","successive cancellation decoding","CRC-aided decoding","deep learning for error correction","AWGN channel"],"falsifier":"Recompute the BER and BLER curves for the SMART variants at their true information rate, namely k minus r information bits carried in 256 channel uses, and compare against SC decoding and DeepPolar at identical rates; if the reported gains collapse or reverse, the claimed trade-off elimination is an equal-rate artifact. Repeat the simulations at the 0.08 dB margin to see whether the apparent BLER improvement survives Monte Carlo noise.","tokens_in":8435,"feed_emoji":"📡","tokens_out":11457,"duration_ms":119548,"temperature":0.7,"pith_summary":"The paper tries to establish that the BER-BLER trade-off of DeepPolar codes is not fundamental but can be removed by changing the decoder and training objective. DeepPolar+ keeps the DeepPolar encoder and successive-cancellation-style sequential decoding, but inserts multi-head self-attention into each kernel decoder and adds a differentiable block-error term to the loss. Around this, the SMART (SNR-matched redundancy) decoder runs several models trained at different SNRs in parallel and uses CRC verification to pick a valid output. For a (256,37) code on AWGN, the authors report a 0.4 dB BER gain over SC decoding at $10^{-4}$ while slightly beating SC on BLER, with DP+SMART adding up to 0.5 dB over SC in BER and 0.6 dB over DeepPolar in BLER at $10^{-3}$. A sympathetic reader would care because it points toward learned codes that do not sacrifice block reliability for bit accuracy, with a 75% reduction in training epochs.","feed_headline":"Neural polar code claims to end the BER-BLER trade-off","feed_subtitle":"Attention decoding plus CRC-guided SNR-matched models beats SC in both error metrics on a (256,37) code.","key_machinery":"The machinery is an attention-augmented neural successive-cancellation decoder paired with a block-aware loss and an SNR-specialized ensemble. Each kernel component network embeds the channel LLRs and previously decoded bits, applies multi-head self-attention with four heads of dimension 32 inside a residual, layer-normalized block, and then continues with feed-forward layers. The total loss is $L_{total} = L_{BCE} + L_{block}$, where $L_{block}$ is the average log-probability of the full codeword, making whole-block mistakes directly costly during training. The SMART decoder encodes a CRC with each message, decodes with M models trained at different SNR pairs, and returns the first candidate whose CRC check passes, falling back to a baseline model otherwise.","core_discovery":"The central claim is that the BER-BLER trade-off observed in DeepPolar codes is an artifact of the decoder architecture and the training loss, not a property of neural polar codes. The authors show that inserting self-attention into the kernel decoders, adding a soft block-error term to the binary cross-entropy objective, and selecting among SNR-specialized decoders with CRC yields simultaneous improvement in both metrics for the (256,37) code over AWGN. They also report that the learned codewords have a pairwise-distance distribution close to that of a random Gaussian codebook, suggesting the model is using the available signal space well, and that training converges in 500 epochs versus the 2000 needed by DeepPolar.","pith_inferences":["At equal information rate, counting the CRC bits as overhead, the SMART gains may shrink; the paper never states the effective payload length, so a fair-rate comparison is the natural experiment to run next.","CRC-guided selection is functionally close to CRC-aided successive cancellation list decoding, so part of the SMART gain may come from the CRC redundancy itself; an ablation that gives the same CRC to the SC baseline would separate the effects.","The reported 0.08 dB BLER improvement over SC sits inside typical Monte Carlo noise at $10^{-3}$ without error bars, so independent replication with fixed codeword budgets would test whether the trade-off is truly eliminated."],"forward_implications":["If the claim holds, neural polar codes can match or exceed conventional SC decoding on block reliability while keeping their bit-level advantage, which is the property needed for practical use.","The SMART ensemble shows that training at several SNRs and picking a valid candidate with CRC can remove the degradation a single learned decoder suffers outside its training SNR.","The 500-epoch convergence result means the attention mechanism and structured loss cut training cost by about 75%, lowering the barrier to deploying learned codes.","Because decoding remains sequential, extending the scheme to list decoding by keeping parallel paths and CRC selection is a direct next step that the authors explicitly identify."],"supporting_citations":[{"why":"Supplies the DeepPolar encoder-decoder structure that DeepPolar+ extends and the main baseline it must beat.","marker":"[30]"},{"why":"Provides the multi-head self-attention operation inserted into each kernel decoder.","marker":"[31]"},{"why":"Introduces CRC-aided decoding, the reliability check the SMART ensemble uses to select a valid output.","marker":"[9]"},{"why":"Defines polar codes and successive cancellation, the conventional decoding baseline compared in the error-rate figures.","marker":"[5]"},{"why":"Provides the KO code baseline against which DeepPolar+ and DP+SMART are compared.","marker":"[23]"},{"why":"Provides the DKO code baseline and the earlier dense-connectivity method by the same authors.","marker":"[25]"},{"why":"Provides Reed-Muller codes with Dumer decoding as a classical non-polar comparison point.","marker":"[1]"}],"fun_headline_variants":["DeepPolar+ ends BER-BLER trade-off","Self-attention and SMART decoding beat polar benchmarks","Neural polar code cracks BER-BLER trade-off with attention","DeepPolar+ improves both error rates on (256,37) code"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes every system sends the same number of information bits per codeword, but the SMART scheme adds CRC bits before encoding and the paper never states the resulting effective rate.","fun_headline_variants_meta":{"raw":{"variants":["DeepPolar+ ends BER-BLER trade-off","Self-attention and SMART decoding beat polar benchmarks","Neural polar code cracks BER-BLER trade-off with attention","DeepPolar+ improves both error rates on (256,37) code"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000408,"raw_usage":{"total_tokens":2142,"prompt_tokens":992,"completion_tokens":1150,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":1081}},"tokens_in":608,"tokens_out":1150,"duration_ms":10646,"temperature":1.0,"reasoning_tokens":1081,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:33:12.062367+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the BER and BLER curves for the SMART variants at their true information rate, namely k minus r information bits carried in 256 channel uses, and compare against SC decoding and DeepPolar at identical rates; if the reported gains collapse or reverse, the claimed trade-off elimination is an equal-rate artifact. Repeat the simulations at the 0.08 dB margin to see whether the apparent BLER improvement survives Monte Carlo noise.","supporting_citations":[{"cited_title":"CRC-aided decoding of polar codes,","cited_arxiv_id":null,"evidence_quote":"Introduces CRC-aided decoding, the reliability check the SMART ensemble uses to select a valid output."}],"review_version":1}