{"id":"a1e39ddc-bf05-422c-9b62-881aa54636ab","arxiv_id":"2506.10121","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Hierarchical pretraining and progressive unfreezing let Kronecker-operation neural codes outperform Reed-Muller codes at orders r=3 and r=4 (block lengths 256 and 512), the first reported extension of KO codes beyond second order.","lead":"This paper trains a new family of neural error-correcting codes, called HiKO, by breaking a large Reed-Muller-style code into smaller pretrained pieces and then fine-tuning the combined network. The resulting codes beat classical Reed-Muller codes at the higher code rates where earlier neural KO codes had failed.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RM baseline curves are unverifiable: the Dumer decoder implementation is unspecified, so the claimed 0.1–0.3 dB HiKO gains over RM(8/9, r=3/4) may reflect a weak comparison rather than a code improvement.","rationale":"The central claim is purely empirical: HiKO codes outperform RM codes at r=3,4. Figure 3 is the only evidence, but the RM curves are not reproducible from the text because Section IV-A omits the decoder implementation and the SNR/power normalization. The reader's concern about Monte Carlo variance is secondary; with 10^7 bits per point the error counts at BER ~1e-4 are roughly 1000-2500, so sampling error is not the main threat. The primary threat is decoder strength and a possible modulation mismatch, both of which are implied by the manuscript's silence on the RM baseline and on the real-valued nature of HiKO. The hierarchical architecture, parameter sharing, and progressive unfreezing are internally consistent, and the code is shared, which is a positive. The CONDITIONAL verdict is appropriate because the main comparison can be settled by a small audit of the baseline script and normalization.","tokens_in":9141,"tokens_out":13064,"duration_ms":158157,"concrete_test":"Reproduce the RM(9,3) and RM(8,4) baseline curves using the soft-decision Dumer decoder exactly as implemented in the public KO-codes repository (including any list size), then recompute the SNR difference at BER=1e-4; if the gap falls within Monte Carlo error, the central claim is not established. Also verify that HiKO codewords and RM BPSK symbols have identical average energy; if not, renormalize and repeat the comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A specifies the HiKO training protocol (batch size, training SNRs, epochs) but says nothing about how the RM comparison curves in Figure 3 are produced beyond the label 'classical Reed-Muller codes.' Section II-B describes Dumer's recursive decoder in equations (3)-(5), but a mathematical description is not an implementation specification: the algorithm has variants (soft vs hard decisions, list size, exact vs approximate recursion), and plain Dumer decoding is known to be suboptimal for high-order RM codes. If the baseline uses a weaker variant, the reported gains of 0.1–0.3 dB at BER 1e-4 (and the 9.98e-5 vs 2.48e-4 point for (9,3) at -1 dB) may be artifacts. A second comparability issue is that the HiKO encoder is real-valued (Section IV-C), while RM is BPSK, and the paper never states the average power constraint or the SNR definition; without a common normalization, the BER comparison is not a clean code comparison. The RM baseline is thus the load-bearing assumption, and it is the least supported part of the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces HiKO, a hierarchical training framework for Kronecker Operation (KO) codes that targets high-order, high-rate configurations (r=3 and r=4) with block lengths 256 and 512. The framework trains smaller constituent KO codes, transfers their parameters into a larger Plotkin-structured encoder/decoder, and then fine-tunes with progressive unfreezing, dropout, and learnable skip connections. The authors report that HiKO codes outperform classical Reed-Muller codes at BER levels around 1e-4, with claimed gains of 0.1 to 0.3 dB, and they present a distance-distribution analysis suggesting that HiKO codes resemble random Gaussian codebooks. The paper also provides a public code repository.","tokens_in":9470,"tokens_out":3801,"duration_ms":48114,"significance":"If the reported comparisons are reliable, this is a meaningful step: it is, to my knowledge, the first claim of KO codes working beyond second order, and it proposes a concrete training strategy (hierarchical parameter transfer plus progressive unfreezing) that could be useful for other neural code families. The strengths are that the HiKO training protocol is described in some detail (Section IV-A), the code is released, and the central claim is an empirical comparison against an external family, not a circular derivation from fitted constants. However, the significance is conditional: the Reed-Muller baseline decoder is not specified, the power/SNR normalization is not stated, and no conventional KO baseline at the target configurations is shown. These gaps directly affect whether the headline 'consistently outperform RM codes' is established, so the contribution, while potentially valuable, is not yet fully supported.","major_comments":[{"comment":"The Reed-Muller baseline is not specified. Section IV-A details the HiKO training protocol (batch size, training SNRs, epochs), but says nothing about how the RM comparison curves in Figure 3 are generated beyond the label 'classical Reed-Muller codes.' Section II-B gives the equations of Dumer's recursive decoder, but a mathematical description is not an implementation specification: the algorithm has variants (soft vs. hard decisions, list size, exact vs. approximate recursion), and Dumer decoding is known to have order-dependent behavior for higher-order RM codes. Without stating the exact decoder, its hyperparameters, and whether it was run under comparable computational conditions, the reported 0.1-0.3 dB gains and the specific point 9.98e-5 vs. 2.48e-4 for (9,3) at -1 dB may reflect a weak baseline rather than a code improvement. This is load-bearing for the central claim and must be fixed by reporting the RM decoder implementation and, ideally, comparing against more than one RM decoding variant.","section":"Section IV-A and Figure 3"},{"comment":"The power constraint and SNR definition are not stated. The HiKO encoder produces real-valued outputs (Section IV-C discusses a quantized version only in the distance analysis), while the RM baseline is BPSK. The paper never specifies the average power constraint (e.g., E[||x||^2] = n) or whether the horizontal axis is Eb/N0 or Es/N0. Without a common normalization, the BER comparison in Figure 3 is not a clean code comparison; a real-valued code can appear better simply because it uses more energy per transmitted symbol. The relation between the training noise variances sigma^2_enc and sigma^2_dec and the test SNR axis should also be clarified.","section":"Section IV-A and Figure 3"},{"comment":"No conventional KO baseline at the target configurations is shown. The introduction says standard KO codes degrade at higher rates, and Figure 1 illustrates two initial KO variants for KO(3,8), but the paper does not compare HiKO(8,3), HiKO(8,4), HiKO(9,3), or HiKO(9,4) against conventionally trained KO codes of the same size and order. Without such a comparison, the reported improvements cannot be attributed to the hierarchical training framework rather than to any end-to-end neural training with the enhanced architecture. This also weakens the 'first successful extension beyond second order' claim, since the reader cannot see whether a plain KO code at r=3 or r=4 indeed fails or how much the hierarchical part contributes.","section":"Section I-C and Section III"},{"comment":"No statistical uncertainty or multiple-run results are reported. The paper states that 10^7 bits are used per SNR point, which for the cited BER values around 1e-4 gives small binomial standard errors (roughly 1e-5 or less), so the pointwise differences are not obviously noise. However, 'consistently outperform' across configurations is a claim about training robustness as well as decoding performance, and no confidence intervals, multiple random seeds, or training-run variance are given. This is secondary to the baseline issue, but it should be addressed with seed variation or a statement that the displayed curves are representative.","section":"Section IV-A and IV-B"}],"minor_comments":[{"comment":"The notation is inconsistent: Figure 1 labels the codes RM(3,8) and KO(3,8), while the rest of the paper uses RM(m,r) and KO(m,r) with m as the length exponent, so these should be RM(8,3) and KO(8,3).","section":"Figure 1"},{"comment":"The LLR definition writes P[yi|xi=0] and P[yi|xi=1], but yi is real-valued for the AWGN channel; the notation should use probability densities (e.g., p(yi|xi)) to be precise.","section":"Equation (3)"},{"comment":"The quantized version of HiKO(8,3) in Figure 4 is not described: no quantization resolution, mapping to binary symbols, or resulting rate/power constraint is given, so the claim of 'binary transmission compatibility' is not supported by the presented details.","section":"Section IV-C"},{"comment":"The parameter-count formula uses 'l' as the input codeword length, but at internal Plotkin nodes the dimension is half of the current block length; the notation should be clarified to avoid confusion between the full code length and the node input size.","section":"Section IV-D"},{"comment":"The cyclical learning-rate scheduler S(eta_max, t) is described verbally, but the cycle period is not stated; without it, the schedule is not fully reproducible.","section":"Equation (17)"},{"comment":"The bibliography entry for the BCH code reference is incomplete: it lacks author names and a citation key, making it difficult to verify.","section":"Reference [2]"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a credible empirical extension of KO codes to rates r=3 and r=4, with a training recipe (hierarchical pretraining, dropout, learnable skip connections, progressive unfreezing) that makes sense and is honestly reported. The numbers look plausible, but the comparison against RM codes rests on an RM baseline whose implementation is never specified. Until that is pinned down, the 0.1-0.3 dB gains are a claim to verify, not a settled result.\n\nWhat is genuinely new: the hierarchical pretraining, where constituent KO(4,3) and KO(4,2) code weights are mapped into larger Plotkin trees and then progressively unfrozen. The paper also shows two earlier attempts that underperformed, which is more transparent than most of this literature. The code is on GitHub, the training protocol is described in enough detail to reproduce, and the distance-distribution analysis is a nice, if non-essential, addition.\n\nWhere it gets soft: the RM decoder is never pinned down. Section II-B gives Dumer's recursion in equations, but that is not an implementation. Dumer decoding has variants (soft/hard decisions, list size, recursion depth), and for high-order RM codes a plain implementation can be substantially suboptimal. If the baseline is weaker than state-of-the-art RM decoding, the reported gains are partly artifacts. Related to that, the HiKO encoder outputs real-valued symbols while RM is BPSK; the paper never states the average power constraint or SNR definition. Without a common normalization, the BER comparison is not a clean code-vs-code comparison. The absence of confidence intervals or repeated runs is a minor but real gap: at 10^7 bits, the difference between 9.98e-5 and 2.48e-4 is statistically solid, but other points, like 2.85e-4 vs 3.49e-4, are only about 20% apart and might move with a different random seed.\n\nThe novelty claim \"first beyond second order\" should be checked against the authors' own Dense KO codes [22]; the paper does not clarify the relationship. That is a citation-pattern issue, not a correctness issue.\n\nBottom line: the central claim is plausible and the training recipe is transferable. The missing baseline details and normalization are exactly what a referee should ask for. This deserves peer review, not a desk reject. My recommendation: send it out, and require the authors to specify the RM decoder, add error bars, and state the power constraint.","headline":"Plausible first extension of KO codes to r=3/4 with a transferable training recipe, but the RM baseline and power normalization need to be nailed down before the gains are trusted.","tokens_in":9920,"tokens_out":4249,"would_cite":true,"duration_ms":47080,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94B35","94B05","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"HiKO trains neural Kronecker-operation codes to beat Reed-Muller at orders 3 and 4.","keywords":["neural channel coding","Kronecker operation codes","Reed-Muller codes","hierarchical training","progressive unfreezing","Plotkin construction","AWGN channel","BER performance"],"falsifier":"Run the identical HiKO training pipeline against an independently implemented RM decoder with the same $10^7$ bits per SNR point and report binomial confidence intervals; if the intervals at the claimed operating points overlap, or if a stronger RM decoder closes the 0.1--0.3 dB gap, the reported consistent gains would not survive. A minimal check: at $-1$ dB, compare the standard error of the reported $9.98\\times10^{-5}$ and $2.48\\times10^{-4}$ BER values for HiKO(9,3) and RM(9,3).","tokens_in":1663,"feed_emoji":"📡","tokens_out":7926,"duration_ms":115476,"temperature":0.7,"pith_summary":"The paper claims that neural error-correcting codes built on the Kronecker-operation (KO) architecture can be trained beyond second order, where standard KO codes degrade, by decomposing a high-rate code into simpler constituent codes and transferring learned parameters from those parts. The authors introduce HiKO, a three-phase training protocol: train small KO components, map their parameters into the larger Plotkin tree, then progressively unfreeze the components while fine-tuning with adaptive learning rates. On the AWGN channel, HiKO codes are reported to outperform classical Reed-Muller codes of the same length and dimension for orders $r=3$ and $r=4$, with gains of 0.1 to 0.3 dB at a bit error rate of $10^{-4}$. If the comparison holds, this would be the first demonstration that learned KO-style codes work at high rates, a regime relevant to practical high-throughput links.","feed_headline":"Neural KO codes beat Reed-Muller at high rates","feed_subtitle":"Reported 0.1-0.3 dB gains over Reed-Muller at BER 1e-4 for block lengths 256 and 512.","key_machinery":"The central object is the hierarchical Plotkin tree shared by RM and KO codes. The paper's mechanism is parameter transfer within that tree: a target $\\mathrm{HiKO}(m,r)$ code reuses the trained encoder and decoder networks of constituent $\\mathrm{KO}(m-1,r)$ and $\\mathrm{KO}(m-1,r-1)$ codes, maps those weights onto the matching subtrees, freezes them, and then unfreezes level by level. The only architectural change to the per-node KO cell is a learnable scalar $\\alpha_i$ multiplying the classical Plotkin term $u \\oplus v$ in the encoder plus dropout and SELU regularized networks in the decoder; this keeps the parameter count essentially unchanged while making the optimization tractable at rates $r>2$.","core_discovery":"HiKO shows that the obstacle to high-rate KO codes is training, not architecture: the Plotkin tree of $\\mathrm{RM}(m,r)$ can be initialized from pre-trained $\\mathrm{KO}(m-1,r)$ and $\\mathrm{KO}(m-1,r-1)$ constituents, and then jointly optimized with dropout, learnable skip connections (realized as one extra scalar per node), and a progressive unfreezing schedule. With this recipe, HiKO(8,3), HiKO(8,4), HiKO(9,3), and HiKO(9,4) all report lower bit error rates than their Reed-Muller counterparts across the tested SNR range, with the largest margin at long block length: at $-1$ dB, HiKO(9,3) reports BER $9.98\\times10^{-5}$ versus $2.48\\times10^{-4}$ for RM(9,3). The paper further reports that HiKO codeword distance distributions resemble random Gaussian codebooks more than the concentrated profile of algebraic RM codes, which it reads as evidence that the learned codes approach Shannon-optimal behavior while retaining recursive decoding.","pith_inferences":["Inference: The hierarchical recipe is not tied to RM or KO codes; any code family whose generator has a recursive Plotkin-like decomposition could receive the same pretraining-then-unfreezing treatment, so the method may transfer to polar or algebraic-geometry constructions.","Inference: A natural stress test is to compare HiKO against RM codes decoded by a stronger soft-decision decoder rather than Dumer's recursive decoder, since the claimed margin is small and could shrink against a stronger baseline.","Inference: The reported distance analysis is geometric, not information-theoretic; proving that the learned encoder-decoder pair actually approaches Gaussian-codebook capacity would require computing achievable rates or mutual information, which the paper does not do.","Inference: A testable extension is to measure the same hierarchical protocol on channels with higher-order modulation or fading, where the Gaussian-codebook approximation is less obviously optimal."],"forward_implications":["If the reported BER curves are correct, HiKO codes give the first learned KO-style codes that beat RM codes at orders 3 and 4, extending neural channel coding to rates previously out of reach.","The 0.1--0.3 dB improvement at BER $10^{-4}$ for block lengths 256 and 512 would be directly usable in short-to-medium block-length regimes where classical codes still compete.","Because the parameter count per node is nearly the same as standard KO codes, the gain costs almost no extra decode complexity at inference time.","The Gaussian-like codeword distance distribution suggests that hierarchical initialization steers learned codes toward random-codebook-like behavior, the classical route to capacity, which the paper argues explains the gains.","The recursive mapping principle stated for HiKO(5,3) extends to the larger codes studied and in principle to even deeper hierarchies, so the framework is not limited to the four reported configurations."],"supporting_citations":[{"why":"Defines KO codes, the neural generalization of the Plotkin construction whose high-rate failure HiKO is designed to fix.","marker":"[21]"},{"why":"Dumer's recursive decoder is the RM decoding algorithm that serves as the classical baseline in the BER comparisons.","marker":"[27]"},{"why":"The Plotkin construction is the recursive building block shared by RM, KO, and HiKO trees.","marker":"[26]"},{"why":"Reed's RM code definition supplies the code family and parameters against which HiKO is compared.","marker":"[1]"},{"why":"Prior KO-architecture work by the same authors that motivates the enhanced network components and the base KO architecture.","marker":"[22]"}],"fun_headline_variants":["HiKO trains KO codes past second order to beat RM","Beyond second order: KO codes outperform Reed-Muller","Training unlocks high-rate KO codes, beating RM","KO codes crack RM at rates 3 and 4 with HiKO","Hierarchical training pushes KO codes past second order"],"cache_read_input_tokens":12160,"weakest_assumption_plain":"The comparison assumes the Reed-Muller baselines are produced by a fair and correctly tuned implementation of Dumer's recursive decoder under the same test conditions, and that the reported BER values are stable enough to support margins of 0.1 to 0.3 dB.","fun_headline_variants_meta":{"raw":{"variants":["HiKO trains KO codes past second order to beat RM","Beyond second order: KO codes outperform Reed-Muller","Training unlocks high-rate KO codes, beating RM","KO codes crack RM at rates 3 and 4 with HiKO","Hierarchical training pushes KO codes past second order"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1392,"prompt_tokens":1009,"completion_tokens":383,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":304}},"tokens_in":625,"tokens_out":383,"duration_ms":5114,"temperature":1.0,"reasoning_tokens":304,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:33:49.171372+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical HiKO training pipeline against an independently implemented RM decoder with the same $10^7$ bits per SNR point and report binomial confidence intervals; if the intervals at the claimed operating points overlap, or if a stronger RM decoder closes the 0.1--0.3 dB gap, the reported consistent gains would not survive. A minimal check: at $-1$ dB, compare the standard error of the reported $9.98\\times10^{-5}$ and $2.48\\times10^{-4}$ BER values for HiKO(9,3) and RM(9,3).","supporting_citations":[{"cited_title":"KO codes: inventing nonlinear encoding and decoding for reliable wireless communication via deep-learning,","cited_arxiv_id":null,"evidence_quote":"Defines KO codes, the neural generalization of the Plotkin construction whose high-rate failure HiKO is designed to fix."},{"cited_title":"Recursive decoding and its performance for low-rate reed- Muller codes,","cited_arxiv_id":null,"evidence_quote":"Dumer's recursive decoder is the RM decoding algorithm that serves as the classical baseline in the BER comparisons."},{"cited_title":"Dense KO codes: Faster convergence and reduced complexity through dense connectivity,","cited_arxiv_id":null,"evidence_quote":"Prior KO-architecture work by the same authors that motivates the enhanced network components and the base KO architecture."}],"review_version":1}