{"id":"8c0e2e51-931a-4fc1-a23c-a8acd03f8d8f","arxiv_id":"2411.18153","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An auto-encoder trained jointly over multiple puncturing patterns learns rate-compatible linear block codes that outperform BCH and 5G LDPC baselines at short block lengths with a 5-iteration belief-propagation decoder.","lead":"This paper trains a single neural auto-encoder to design linear block codes that support several code rates by puncturing, sharing one set of parameters across rates. A generalist reader might care because rate-compatible coding is central to adaptive wireless links, and storing one model instead of many could simplify 5G and 6G transceivers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 5-iteration BP comparison is not complexity-fair: the learned parity-check matrix is dense, so per-iteration message cost is far above that of BG2 LDPC, and the claimed BER margins may not survive equal-complexity accounting.","rationale":"The reader's weakest assumption focuses on the fixed truncation puncturing pattern, which the authors explicitly acknowledge as a limitation and which limits generalizability rather than threatening the validity of the reported comparison. My concern is more central to the stated comparison: the paper compares BER under equal iteration count but not equal decoding complexity. The learned parity-check matrices are not constrained to be sparse, so the per-iteration BP cost is likely much higher than for the sparse BG2 LDPC baselines, and BCH parity-check matrices are also dense. If the claimed BER advantage disappears when total message updates are equalized, then the practical value of a single parameter set replacing multiple rate-compatible codes is not established. This does not change the appropriate verdict, which remains conditional: the authors should provide complexity-normalized comparisons or release code and matrices so that edge counts and equal-cost BER curves can be checked. I therefore keep the reader's conditional verdict unchanged, while sharpening the specific condition that should be met before the performance claim is accepted.","tokens_in":8743,"tokens_out":13093,"duration_ms":120237,"concrete_test":"Extract the learned H for both AI-RC-LBC families and count the number of active edges in each rate's submatrix. Then rerun the Section IV-A experiments under a fixed message-update budget instead of fixed 5 iterations: for each code, choose the number of BP iterations so that total edge updates are equal (iterations = budget / edges), and plot BER versus SNR at equal budget, and also plot BER versus edge-update count at a fixed SNR. If the AI-RC-LBC margin over BG2 LDPC and BCH shrinks below the reported 2-3 dB or reverses, the headline claim is an artifact of unequal per-iteration complexity; if the margin persists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A claims that AI-RC-LBC outperforms BCH and BG2 LDPC codes in all compared cases, with all decoders set to 5 iterations. The load-bearing assumption is that equal iteration count is a fair proxy for equal decoding cost. That assumption is insecure: the learned H is generated by the Matrix-Gen module in Section III-B with no sparsity constraint. H1 is a fully learnable (n-k) by k block, and H2 is lower triangular with every subdiagonal entry learnable, so the resulting parity-check matrix is dense with O(n(n-k)) edges. BP per-iteration cost is proportional to the number of edges, so 5 iterations on the learned code can require far more message updates than 5 iterations on the sparse BG2 parity-check matrices. The paper reports no density or edge-count comparison and no complexity-normalized BER curves. Thus the central practical claim that one learned code can replace conventional rate-compatible codes is not established: the reported margin may reflect additional computation rather than a better code/decoder pair. This concern is distinct from the acknowledged fixed puncturing pattern; even taking the truncation pattern as given, the decoder-complexity comparison must be made fair before the headline claim can be accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an auto-encoder (AE) framework for learning rate-compatible linear block codes (RC-LBCs). A Matrix-Gen module learns a low-rate precode's parity-check matrix H and derived generator matrix G; higher code rates are obtained by puncturing trailing parity bits. The decoder is a rate-compatible neural belief propagation (RC-NBP) network whose edge weights are jointly trained with the code under a multi-task objective across rates. Experiments compare the learned codes with BCH and 5G NR LDPC BG2 baselines under 5-iteration BP, with single-rate AI-based LBCs, and across systematic versus non-systematic H. The authors report BER gains in all compared cases and claim that one learned parameter set can replace multiple single-rate conventional codes.","tokens_in":9021,"tokens_out":4923,"duration_ms":47106,"significance":"If the empirical claims hold, the work is a useful step toward rate-compatible code design for short block lengths, where conventional LDPC codes are known to be weak, and it offers the practical advantage of storing a single model for multiple rates. The multi-task parameter-sharing scheme is a sensible way to exploit the nested structure of punctured codes, and the comparison against separately trained single-rate AI-LBCs helps isolate the benefit of multi-rate training. However, the paper ships no code or data, the BER curves have no error bars, and the headline comparison is not complexity-normalized; these gaps currently prevent the practical advantage from being established. The fixed truncation puncturing pattern is explicitly acknowledged as a limitation, which is appropriate, but it also means the claimed gains are demonstrated only for that pattern.","major_comments":[{"comment":"The 'fair comparison' in Section IV-A fixes all decoders to 5 iterations, but the learned parity-check matrices are dense: H1 is fully learnable and H2 is lower triangular with every subdiagonal entry learnable (Section III-B, Eq. (3)), so the per-iteration BP message cost is roughly O(m n) with m = n-k, whereas the 5G BG2 LDPC baselines are explicitly designed to be sparse. The paper reports no edge counts, no density comparison, and no complexity-normalized BER curves. The claim that AI-RC-LBC 'outperforms' the baselines is therefore not yet established as a coding/decoding gain rather than an artifact of higher per-iteration complexity. Please report edge counts and BER versus message-update cost or an equivalent complexity-normalized comparison.","section":"Section IV-A; Section III-B"},{"comment":"The BCH baselines are evaluated with a belief-propagation decoder ('conventional BP decoding algorithm', Section IV-A), not with the algebraic decoder normally used for BCH codes. Because BCH codes are not designed for BP decoding, this choice can only weaken the baseline; the headline statement that the learned code 'outperforms BCH codes' should either be qualified as 'under BP decoding' or repeated with the standard algebraic BCH decoder. In addition, the ML-BCH curves in Fig. 4 are not described: it is unclear how they are computed and whether they are intended as lower bounds on the BP results.","section":"Section IV-A"},{"comment":"The empirical claims are supported by single BER curves with no error bars, no number of random seeds, and no numerical tables, and the manuscript includes no code or data release. For a learned-code paper, where two training runs can differ by fractions of a dB, the statement that the learned code 'outperforms ... in all cases' needs confidence intervals or repeated-run statistics; otherwise the shrinking margins at the higher code rates may not be statistically meaningful.","section":"Section IV-A; Figs. 4-6"}],"minor_comments":[{"comment":"Appendix A and Appendix B contain placeholder text ('Appendix one text goes here', 'Appendix two text goes here') and the acknowledgment section is incomplete; these must be completed before the manuscript can be considered a finished submission.","section":"Appendix A; Appendix B; Acknowledgment"},{"comment":"The footnote contains an apparent typo: '(21,31) BCH codes' should presumably read '(31,21) BCH codes'.","section":"Section IV-A, footnote 1"},{"comment":"There is a duplicated phrase: 'tries to learn learns optimal weights' should be 'tries to learn optimal weights'.","section":"Section II-C"},{"comment":"The word 'empolyed' is a typo for 'employed'.","section":"Section III-B, Eq. (4)"},{"comment":"The text says the decoder has learnable weights and biases in Eq. (2), but then states that 'we only adapt weight multipliers in our experiments'; please clarify whether the biases beta are trained or fixed, and whether the same setting is used for the RC-NBP decoder.","section":"Section III-C"},{"comment":"The Fig. 7 legend labels the non-systematic curves as (31,16), (21,16), and (16,11), while the text describes the precode as C(31,11); the labels appear inconsistent and should be corrected.","section":"Section IV-C, Fig. 7"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as an incomplete draft: placeholder appendices, an incomplete acknowledgment, and several typos. The editor may wish to treat this as an early preprint rather than a submission-ready article. The main technical concern, complexity fairness of the 5-iteration comparison, is fixable with additional experiments, so I am not recommending rejection; however, the paper should not be accepted without complexity-normalized BER comparisons, clarification of the BCH decoder, and reproducibility details such as code/data availability or repeated-run statistics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you care about learned short-block codes. The new thing is modest but real: instead of learning one code per rate, they train one auto-encoder whose single parameter set yields a rate-compatible family by puncturing the last parity bits, sharing parameters across rates through multi-task training. That combination appears to be new relative to Larue et al. [14] and the authors' own RC-NBP decoder [16]. The non-systematic extension is also a small plus. The paper deserves credit for stating what it does not do: the puncturing pattern is not optimized, and it says so in Section I.\n\nWhat is not established is the headline BER claim. The evaluation compares 5 iterations of BP on the learned codes against 5 iterations on BCH and BG2 LDPC baselines, and calls that fair. It is not complexity-fair. The learned parity-check matrix is dense: H1 is fully learnable and H2 is lower triangular with all subdiagonal entries learnable. Per-iteration BP cost scales with the number of edges, so five learned-code iterations can cost many times more message updates than five iterations on sparse BG2. The paper reports no edge counts and no complexity-normalized curves. The margin over LDPC at 2-3 dB may be purchased with extra computation rather than a genuinely better code/decoder pair. That is the load-bearing weakness, and it survives even if you accept the fixed truncation puncturing pattern.\n\nOther soft spots, in proportion: no code or data, no error bars, placeholder appendices, and a cut-off acknowledgment. Evaluating BCH under BP is fair only if the claim is that the learned code works with a simple decoder, but BCH is not designed for 5-iteration BP, and the paper's own ML-BCH curve shows that a stronger decoder flips the comparison. The puncturing limitation is acknowledged, so treat it as a limitation, not a hidden flaw.\n\nFor a subfield paper, the engineering integration is useful, but the artifact and the complexity question matter. I would not cite it until the authors release code and complexity-fair BER curves. It does deserve a serious referee, though: the idea is coherent, the empirical claim is falsifiable, and the missing pieces are fixable in revision.","headline":"Useful multi-task rate-compatible learned-code idea, but the headline BER margin is not complexity-fair against sparse LDPC baselines, and the manuscript is incomplete.","tokens_in":9545,"tokens_out":2025,"would_cite":false,"duration_ms":19778,"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":"One auto-encoder learns a family of rate-compatible codes that beat BCH and LDPC baselines.","keywords":["rate-compatible codes","linear block codes","auto-encoder","belief propagation","neural belief propagation","multi-task learning","puncturing","short block lengths"],"falsifier":"Train the same auto-encoder but replace the end-bit truncation with optimally chosen puncturing patterns, for example by density evolution or exhaustive search, and compare BER under five BP iterations against BCH(31,11) and BG2(100,20); if no puncturing pattern can reproduce the reported gains, the rate-compatibility claim as stated fails.","tokens_in":8496,"feed_emoji":"📡","tokens_out":5219,"duration_ms":43945,"temperature":0.7,"pith_summary":"This paper claims that a single auto-encoder can learn a family of rate-compatible linear block codes, one parameter set that serves several code rates by deleting parity bits before transmission. If true, a transceiver no longer needs to store one model or parity-check matrix per rate; the same learned code adapts to channel conditions through puncturing. The paper reports that these learned codes outperform conventional BCH codes and 5G BG2 LDPC codes at similar rates under belief-propagation decoding with five iterations, and that multi-rate training itself improves each rate relative to training a single-rate code. This matters for short-block 6G-style links where classical codes lose their asymptotic advantage and where adaptive-rate operation is required.","feed_headline":"One auto-encoder learns code rates that beat BCH and LDPC","feed_subtitle":"A single trained parameter set covers multiple rates, saving storage and beating short-length baselines under BP.","key_machinery":"The load-bearing object is the Matrix-Gen module inside the auto-encoder, which turns learnable float weights $W$ into a binary parity-check matrix $H=[H_1|H_2]$ with $H_2$ lower triangular, then derives the systematic generator $G$ through linear transformations; a differentiable step function keeps this binarization trainable. Around this sits the rate-compatible neural belief-propagation decoder, which unfolds BP iterations into cells with per-edge learnable weights and activates only the neurons corresponding to the punctured code rate. Multi-task learning ties the rates together: parameters for a given rate are updated using data from all lower rates, exploiting the nested structure of punctured codewords. The fixed puncturing rule, deleting the last parity bits, is what converts the single learned precode into the family of rates.","core_discovery":"The central claim is that the proposed auto-encoder, called AI-RC-LBC, jointly optimizes the parity-check matrix $H$ and the generator matrix $G$ together with decoder weights, so that puncturing the last parity bits of a low-rate precode yields good codes at every intermediate rate. The paper reports BER curves showing the learned RC-LBC outperforms BCH(31,11), BCH(21,11), BCH(16,11) and BG2 LDPC(100,20) and (60,20) in all tested cases with similar rates under BP decoding with five iterations; at the lowest rate and BER $10^{-4}$ the margin is about 3 dB over BCH and 2 dB over LDPC. It also reports that the multi-task-trained RC-LBC beats separately trained single-rate AI-LBCs at the same rates, and that non-systematic parity-check matrices outperform systematic ones when decoded with BP. The intended conclusion is that rate compatibility is not a compromise but a regularizer: one learned matrix family can replace several single-rate codes with better bit-error rate and lower storage.","pith_inferences":["The paper fixes puncturing to end-bit deletion without optimizing it; a natural extension is to optimize or search puncturing patterns jointly with the matrix, which could widen the reported margins or reveal which patterns the learned matrices prefer.","Because the multi-task training appears to regularize the code, the same parameter-sharing idea could be tested on longer codes or on fading channels, though the exponential codeword count limits direct scaling.","The claim that non-systematic matrices help under BP suggests the learned codes could be combined with decoder-specific objectives, such as minimizing short cycles, to push further at very short block lengths."],"forward_implications":["A single trained parameter set covers multiple code rates, replacing several stored codes and simplifying rate-adaptive systems.","The learned codes work with a conventional BP decoder, so they can be dropped into existing receivers without a neural decoder.","Multi-task training acts as regularization, giving the rate-compatible code better BER than single-rate AI codes at the same rate.","Allowing non-systematic parity-check matrices improves BP decoding performance, especially when the decoder runs few iterations.","The performance gains are largest at the lowest code rate and shrink as puncturing shortens the code, so the scheme is best suited to rate ranges close to the precode."],"supporting_citations":[{"why":"Provides the neural belief propagation auto-encoder baseline and the differentiable step function used to binarize H.","marker":"[14]"},{"why":"Supplies the rate-compatible NBP decoder whose per-edge weights are trained inside the auto-encoder.","marker":"[16]"},{"why":"Defines the 5G BG2 LDPC codes used as performance baselines.","marker":"[7]"},{"why":"Establishes rate-compatible puncturing of LDPC codes, the concept the scheme builds on.","marker":"[2]"},{"why":"Represents the alternative AI-based code design approach that learns codes without rate compatibility.","marker":"[8]"},{"why":"Introduces learnable weights for belief-propagation decoding, the basis of the RC-NBP parameterization.","marker":"[11]"},{"why":"Justifies the multi-task training strategy as regularizing and reducing overfitting.","marker":"[17]"}],"fun_headline_variants":["One auto-encoder yields rate-compatible codes that beat BCH and LDPC","Single trained model replaces multiple codes, beats BCH and LDPC","Rate-compatible codes learned by one AE beat BCH and LDPC","One AE learns multiple rates with better BER than BCH and LDPC","Learned puncturing lets one model beat BCH and LDPC across rates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result depends on the unoptimized choice to build higher-rate codes by simply deleting the last parity bits; if a different puncturing pattern is needed to make rate-compatible codes competitive, the reported margins over BCH and LDPC codes could shrink or reverse.","fun_headline_variants_meta":{"raw":{"variants":["One auto-encoder yields rate-compatible codes that beat BCH and LDPC","Single trained model replaces multiple codes, beats BCH and LDPC","Rate-compatible codes learned by one AE beat BCH and LDPC","One AE learns multiple rates with better BER than BCH and LDPC","Learned puncturing lets one model beat BCH and LDPC across rates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001205,"raw_usage":{"total_tokens":5124,"prompt_tokens":888,"completion_tokens":4236,"prompt_tokens_details":{"cached_tokens":768},"prompt_cache_hit_tokens":768,"prompt_cache_miss_tokens":120,"completion_tokens_details":{"reasoning_tokens":4133}},"tokens_in":120,"tokens_out":4236,"duration_ms":280997,"temperature":1.0,"reasoning_tokens":4133,"cache_read_input_tokens":768,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:28:14.491421+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same auto-encoder but replace the end-bit truncation with optimally chosen puncturing patterns, for example by density evolution or exhaustive search, and compare BER under five BP iterations against BCH(31,11) and BG2(100,20); if no puncturing pattern can reproduce the reported gains, the rate-compatibility claim as stated fails.","supporting_citations":[{"cited_title":"Neural belief propagation auto-encoder for lin ear block code design,","cited_arxiv_id":null,"evidence_quote":"Provides the neural belief propagation auto-encoder baseline and the differentiable step function used to binarize H."},{"cited_title":"Rate compatible LDPC neural decoding network: A multi-task learning approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the rate-compatible NBP decoder whose per-edge weights are trained inside the auto-encoder."},{"cited_title":"Partnership Project (3GPP), Nov","cited_arxiv_id":null,"evidence_quote":"Defines the 5G BG2 LDPC codes used as performance baselines."},{"cited_title":"Rate-compatible punct uring of low-density parity-check codes,","cited_arxiv_id":null,"evidence_quote":"Establishes rate-compatible puncturing of LDPC codes, the concept the scheme builds on."},{"cited_title":"AI coding: Le arning to construct error correction codes,","cited_arxiv_id":null,"evidence_quote":"Represents the alternative AI-based code design approach that learns codes without rate compatibility."},{"cited_title":"Learning to decode protograph LDPC codes,","cited_arxiv_id":null,"evidence_quote":"Introduces learnable weights for belief-propagation decoding, the basis of the RC-NBP parameterization."}],"review_version":1}