{"id":"c7a72f89-d2ab-4a4d-881e-2f199e65d703","arxiv_id":"2412.02053","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A reinforcement-learning agent and a graph-neural-network decoder, trained in alternating rounds, jointly design short error-correcting codes that beat standard LDPC and BCH schemes on simulated AWGN channels.","lead":"This paper jointly trains an AI code designer and a graph-neural-network decoder for short error-correcting codes, using reinforcement learning to search over parity-check matrices and learned edge weights to improve message passing. On simulated noisy channels, the pair beats standard LDPC and BCH schemes at short block lengths, and the decoder transfers to other code sizes without retraining.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reward-proxy risk is the load-bearing soft spot: the BP-BER reward (Eq. 7) already yields H*_d with MHD=3 (Table II), and the paper never shows that the iterative EW-GNN reward used for H*_AE avoids the same overfitting.","rationale":"I read the paper as an empirical demonstration rather than a formal construction, and it deserves credit for reporting the H*_d failure and the (63,45) MLD inferiority; those admissions make the remaining risk precise instead of diffuse. The central claim is not mathematically proved and rests on Monte Carlo curves, so the key question is whether the proxy used to search for codes tracks the quantity advertised. The paper's own Table II is direct evidence that the BP-BER reward can select a code with MHD=3 that is MLD-inferior to the baseline, while still looking superior under BP. The autoencoder's headline gains depend on H*_AE, which is found by the same kind of decoder-BER reward (with EW-GNN in the loop). The paper never gives dmin(H*_AE), its weight spectrum, or an independent verification of the MLD curves used to claim gains over LDPC+MLD, so the 1.28 dB and 0.63 dB gains are not yet tied to an intrinsic code property. This is not a reason to reject: the reported experiments are plausible and the paper is honest about negative results. It is, however, the reason the accept decision must remain conditional on a full characterization of H*_AE and on reproducible MLD baselines, including a best-known (32,16) code such as RM(2,5), which has dmin=8 and would sharpen the comparison. The reader's weakest assumption identifies the same issue, and my read does not move the verdict.","tokens_in":23895,"tokens_out":10471,"duration_ms":106589,"concrete_test":"Release H*_AE and G*_AE for the (32,16) design; compute exact dmin and run exhaustive MLD (2^16 codewords) over 4-8 dB, then compare with CCSDS LDPC, NBP-AE, and with RM(2,5) = extended BCH (32,16,8). Separately, repeat the iterative autoencoder training with reward SNRs of 4, 5, 6, and 7 dB. If dmin(H*_AE) < 4, or if the 1.28 dB MLD gain disappears under any reward SNR, then the claimed 'superior error-correction capabilities' are tied to a single training operating point rather than to the code.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section VI-A/Table II is the internal warning: the BP-BER reward (Eq. 7, simulated at SNR=6 dB with 8-iteration BP) selects H*_d with MHD=3 and MLD-inferior BER to the LDPC baseline (2.64e-4 vs 2.29e-4 at 5 dB), even though H*_d looks good under BP. The paper says the sparse, low-cycle structure favored by the reward comes at the cost of minimum distance. The abstract's general claim, however, is about the autoencoder code H*_AE, which is selected by the same type of decoder-BER reward (with EW-GNN replacing BP in the iterative loop of Section V). What would have to be true for the central claim to hold is that this second reward does not repeat the H*_d failure. That is the load-bearing condition. The paper reports only the number of 4-cycles of H*_AE (77); it never reports dmin(H*_AE), its weight enumerator, or an independently verified MLD BER curve. Thus the 1.28 dB MLD gain over LDPC and the 0.63 dB gain over LDPC+MLD in Figs. 10/12 are not yet connected to an intrinsic code-quality measure; they could be operating-point artifacts of matching EW-GNN's fixed point at the reward SNR. Honest acknowledgements elsewhere (H*_d MLD-inferior, learned (63,45) MLD-inferior to BCH) increase trust but do not close this gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an end-to-end channel encoding/decoding system, termed a GNN-based auto-encoder, for short linear block codes. A DRL-based agent, trained with DDPG, generates parity-check matrices by flipping entries of an initial matrix; the agent can be rewarded either by a simulated BER at a single SNR (Eq. 7) or by a structural reward involving minimum distance and short cycles (Eq. 8). An edge-weighted GNN decoder (EW-GNN) is introduced, which aligns with belief propagation but learns multiplicative edge weights to suppress unreliable messages. The designer and decoder are trained iteratively, alternating between code design and decoder fine-tuning. Experiments report BER gains over CCSDS LDPC and BCH codes, as well as over the NBP-AE baseline, for (32,16) and (63,45) codes, and demonstrate decoder transferability across code lengths and rates without retraining.","tokens_in":24136,"tokens_out":4626,"duration_ms":43015,"significance":"If the reported gains are reproducible, the work advances learning-based code design by combining a scalable GNN decoder with a DRL-based parity-check matrix search, and it explicitly targets the short-block-length regime where BP and LDPC codes are known to be suboptimal. The paper has notable strengths: the evaluation protocol is clearly described (10,000 bit errors per point), the decoder's transferability across code lengths is directly demonstrated (Figs. 8-9), and the authors honestly disclose negative results, including that the BP-BER-rewarded code H*_d is MLD-inferior to LDPC (Table II) and that the learned (63,45) code is MLD-inferior to BCH (Fig. 11). These admissions increase confidence in the empirical protocol. However, the central claim that the auto-encoder significantly surpasses LDPC with MLD depends on the MLD curves for the learned code, whose intrinsic code quality (minimum distance, weight enumerator) is never reported and whose MLD implementation is not described.","major_comments":[{"comment":"The paper's own Table II shows that the BP-BER reward of Eq. (7), evaluated at a single SNR (6 dB) with 8-iteration BP, selects a code H*_d whose minimum distance is 3 and whose MLD BER at 5 dB (2.64e-4) is worse than the CCSDS LDPC baseline (2.29e-4). The auto-encoder code H*_AE, whose headline MLD gains appear in Figs. 10 and 12, is selected by the same type of decoder-BER reward, with EW-GNN replacing BP in the iterative loop of Section V. The paper reports only the number of 4-cycles for H*_AE (77 in Section VI-C1) and never reports dmin, the weight enumerator, or an independently verified MLD curve. Without this information, the claimed 1.28 dB MLD gain over LDPC (Fig. 10) and 0.63 dB gain over LDPC with MLD (Fig. 10) could be operating-point artifacts of matching EW-GNN's fixed point at the reward SNR rather than intrinsic code quality. Please add dmin(H*_AE) and a validation of the MLD results against an independent MLD implementation or known code database, such as [39].","section":"VI-A and Table II"},{"comment":"The MLD curves for H*_AE are obtained without any description of the maximum-likelihood decoding algorithm used. The paper states 'We also test the MLD performance' but does not specify whether this is brute-force nearest-codeword search, trellis-based decoding, or some other exact MLD method, nor does it describe how decoding errors are counted (message bits or codeword bits). This matters because H*_AE is a non-standard code and its MLD curve cannot be cross-checked by the reader without either the parity-check matrix or the decoding algorithm. Please specify the MLD implementation, its complexity, and provide a sanity check by reproducing a known MLD curve (e.g., for the CCSDS LDPC baseline) with the same implementation.","section":"VI-C1 and Fig. 10"},{"comment":"Fig. 13 is used to claim that the coding gain between H*_2 and H*_3 'diminishes' and to attribute this to EW-GNN scalability. However, H*_1 is designed with a BP-based reward, whereas H*_2 and H*_3 are designed with the EW-GNN-based reward; the comparison therefore conflates a change of reward function with the convergence of the iterative training. A more direct convergence test would be to run more than three iterations or to report the reward values and code properties (in particular dmin and the cycle distribution) at each iteration. Please also clarify whether the 'MLD' curves in Fig. 13 are computed with the same decoding algorithm as in Fig. 10 and whether the same training setup was used for all three iterations.","section":"VI-C3 and Fig. 13"}],"minor_comments":[{"comment":"The definition 'SNR = 1/σ_n^2' is incomplete for a BPSK AWGN channel; please state whether this is Es/N0 or Eb/N0 and clarify the normalization of the symbol energy, since the coding gain comparisons in Section VI depend on the SNR convention.","section":"Section III-A"},{"comment":"Figure 3 states that the flipping threshold α_f is 0.5, while Table I reports α_f = 0.3; please align the illustrative example with the actual hyper-parameter setting.","section":"Figure 3 and Table I"},{"comment":"The text says H*_AE has '77 4-step cycles' and that NBP-AE has '172 short cycles'; the metric for NBP-AE is not defined in Table II or elsewhere. Please report the same cycle metric for all compared codes for a fair comparison.","section":"Section VI-C1"},{"comment":"The structural reward r_t^s is not normalized to a scale comparable with the BER reward r_t^d in Eq. (7); please clarify whether the two rewards are ever mixed in one training run and how their relative scales are handled in the DDPG training.","section":"Eq. (8)"},{"comment":"There are minor language issues: 'auto-encoder based' should be hyphenated, and the caption of Fig. 7 contains 'Peformance' instead of 'Performance'.","section":"Abstract and Fig. 7 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper builds directly on the authors' ICC 2023 paper [20], and the overlap should be stated explicitly in the final version. The most serious risk is that the MLD gains claimed for H*_AE are not verified against an independent decoding method or against any intrinsic code property, and the paper's own evidence on H*_d shows that decoder-BER rewards can select codes with poor MLD performance. The authors' candid reporting of negative results is commendable, but the missing dmin and validation for the headline auto-encoder code is the kind of gap that a coding-theory reviewer will see as load-bearing. From a journal-fit perspective, the incremental contribution over [20] plus the iterative joint-training scheme is real, but the verification depth should be raised before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis is a serious attempt at joint code and decoder design, and it is more than a patchwork of known tricks. The integration—a DDPG agent that flips parity-check entries on a lattice graph, an edge-weighted GNN decoder aligned with BP, and an alternating training loop—produces concrete artifacts: the learned matrices H*_s and H*_AE, the EW-GNN decoder, and the Section V protocol. Credit where due: the paper is unusually honest about failures. It reports that H*_d has MHD=3 and is MLD-inferior to the LDPC baseline, and that the learned (63,45) code is MLD-inferior to BCH. That honesty makes the empirical protocol (BER to 10,000 errors per point, Figs. 7-13) more believable.\n\nThe soft spot is exactly where the stress-test note lands. The BP-BER reward (Eq. 7, at 6 dB with 8-iteration BP) selects H*_d, which looks good under BP but has weak minimum distance. The headline auto-encoder gains come from H*_AE, selected by the same type of decoder-BER reward, with EW-GNN replacing BP. The paper never reports dmin(H*_AE), its weight enumerator, or an independently verified MLD curve. So the 1.28 dB MLD gain over LDPC in Fig. 10 could be an operating-point artifact of matching EW-GNN's fixed point at the reward SNR. That is the load-bearing condition, and it is not verified.\n\nOther concerns are real but minor by comparison. The MLD implementation for the designed (63,45) code is unspecified. The baseline set is thin: a single CCSDS LDPC code, no best-known linear codes from Grassl, no polar-CRC. Gains of 0.1-0.2 dB come without error bars or multiple seeds. And the learned matrices and decoder weights are not released, which blocks reproduction of the central curves.\n\nThe paper is not fatally flawed. The EW-GNN decoder alone—trained on a short code and applied to longer codes without retraining—is a solid standalone result. The auto-encoder is worth engaging with, but the abstract overstates what is currently established. This is for people working on neural decoders and learned code construction; the decoder scalability result is the most reusable piece.\n\nSend it to review. A serious referee should demand the artifacts, the dmin and weight enumerator for H*_AE, an explicit MLD description, and stronger baselines. With those, the main claim can actually be tested.","headline":"Serious joint code/decoder design with honest failure reporting, but the headline MLD gains rest on an unverified reward-proxy assumption.","tokens_in":24849,"tokens_out":3117,"would_cite":false,"duration_ms":27614,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94B05","94B35","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Joint machine-learning design of short codes and their decoders beats LDPC and BCH at short block lengths, the paper claims.","keywords":["channel coding","short linear block codes","deep reinforcement learning","graph neural networks","belief propagation","auto-encoder","URLLC","Tanner graph"],"falsifier":"Compute the minimum Hamming distance of the auto-encoder code $H^*_{AE}$ and evaluate its BER at SNRs well below and above 6 dB (e.g., 3 dB and 8 dB) with the EW-GNN decoder and with MLD; if the code's low-density, low-cycle structure makes it worse than the reference LDPC under MLD away from the training point—as already happens for the BER-reward-only code $H^*_d$—then the single-SNR reward is not a valid proxy and the headline gains are operating-point artifacts.","tokens_in":23511,"feed_emoji":"📡","tokens_out":6846,"duration_ms":56299,"temperature":0.7,"pith_summary":"This paper claims that a short linear block code and its decoder can be designed together by machine learning, and that the resulting pair beats conventional coding at the block lengths used in ultra-reliable low-latency communications. The authors build a reinforcement-learning agent that generates parity-check matrices by flipping entries of an initial matrix, and pair it with a graph-neural-network decoder whose per-edge weights suppress unreliable messages from short cycles. Trained jointly in alternation, the code and decoder reach coding gains of about 1 dB over a standard (32,16) LDPC code decoded with belief propagation, 0.63 dB over the same LDPC code decoded with maximum-likelihood decoding at 30 iterations, and 1.81 dB over a (63,45) BCH code with BP. If true, this would mean learned codes can replace hand-designed algebraic codes in short-block regimes while keeping decoding complexity similar to BP.","feed_headline":"Learned code plus GNN decoder beats classic short codes","feed_subtitle":"Jointly designed parity-check matrices and graph decoders gain up to 1.81 dB over LDPC and BCH at short block lengths.","key_machinery":"The argument runs on three coupled objects. First, the DRL code designer treats the parity-check matrix as a lattice graph and each entry as a node; a message-passing matrix neural network (the actor and critic) outputs a flip probability for every entry, and the agent explores the space of full-rank matrices. Its reward is either the simulated BER at 6 dB with an 8-iteration BP decoder, $r_t^d = |\\ln \\epsilon_b(S_{t+1})|$, or a structure reward $r_t^s = d_{\\min}/\\alpha_d + \\alpha_c/(c_4 + \\alpha_c)$ that rewards large minimum Hamming distance and few 4-cycles, with $\\alpha_d = 8$ and $\\alpha_c = 500$. Second, the EW-GNN decoder runs BP-style message passing on the Tanner graph but multiplies each check-to-variable message by a learned edge weight $w^{(t)}_{u_j \\to v_i}=g(|\\mu^{(t)}_{u_j\\to v_i}|, \\delta^{(t)}(\\mu_{u_j\\to v_i}), \\delta^{(t-1)}(\\mu_{v_i\\to u_j}), \\delta^{(t-1)}(h_{v_i})|\\theta_g)$, with the $\\tanh^{-1}$ in the check update clipped for differentiability; because all edges share the same MLP, the parameter count is independent of code length. Third, alternating training lets the decoder's BER become the reward for the next code-search iteration, so code and decoder co-adapt, and the paper reports convergence in about three iterations.","core_discovery":"The central claim is that the proposed auto-encoder—a DRL-based code designer that searches parity-check matrices plus an edge-weighted GNN decoder—produces codes whose error-correction performance surpasses several classical schemes at short block lengths. In the (32,16) case, the learned code with the EW-GNN decoder gains about 1 dB over a standard LDPC code with BP, and gains 0.63 dB over LDPC with maximum-likelihood decoding when the decoder is run for 30 iterations; the learned code also outperforms LDPC under MLD by 1.28 dB. For (63,45), the auto-encoder gains 1.81 dB over BCH with BP. The paper further claims that the EW-GNN decoder, once trained on one linear block code, can decode other codes of different lengths and rates without retraining, and that the iterative joint training converges within about three iterations.","pith_inferences":["The reward is a single-SNR, single-decoder proxy: the paper's own results show that the BER-reward-only code has minimum Hamming distance 3 and is worse than LDPC under MLD. An analogous failure for the auto-encoder code would mean the reported gains are tied to the 6 dB operating point and the 8-iteration BP/EW-GNN decoder, not to intrinsic code quality.","The structure-reward constants ($\\alpha_d=8$, $\\alpha_c=500$, and cycle counting limited to 4-cycles) are hand-picked; tuning or replacing them with differentiable estimates of minimum distance and girth could change the discovered codes, and the paper does not study this sensitivity.","Because EW-GNN transfers across code lengths, the same machinery could be probed on much longer codes or on nonbinary or quantum LDPC families, but the paper only demonstrates transfer up to length 256 at rate 1/2."],"forward_implications":["Short-block URLLC links could use learned codes that outperform classical LDPC and BCH codes under practical iterative decoding, narrowing the gap to finite-length bounds at these block lengths.","A single trained decoder can be reused for different code lengths and rates, so a system could switch codes dynamically without retraining the receiver.","The learned code plus EW-GNN can beat maximum-likelihood decoding of a standard LDPC code at high SNR while keeping BP-like per-iteration complexity, relaxing the usual performance-complexity trade-off.","Joint encoder-decoder co-design converges in a few alternations, suggesting that end-to-end optimization is tractable for practical block lengths rather than only toy sizes."],"supporting_citations":[{"why":"Supplies the standard (32,16) LDPC code used as state initialization and as the main baseline for the auto-encoder comparisons.","marker":"[45]"},{"why":"Provides NBP-AE, the learned auto-encoder baseline the proposed system must beat, and the GNBP decoder compared in the simulation figures.","marker":"[21]"},{"why":"Defines the neural belief propagation decoder used as a benchmark for EW-GNN on BCH and LDPC codes.","marker":"[15]"},{"why":"Introduces the GNN decoder over Tanner graphs that EW-GNN extends with edge weights.","marker":"[19]"},{"why":"Prior work by the authors establishing the scalable GNN decoding structure whose parameter count does not change with code length.","marker":"[20]"},{"why":"Provides the deep deterministic policy gradient algorithm underlying the DRL code designer.","marker":"[32]"},{"why":"Supplies the message-passing neural network formalism used by both the MPMNN actor-critic and the EW-GNN decoder.","marker":"[36]"},{"why":"Provides the database of codes used for state initialization and the BCH codes used in the (63,45) comparisons.","marker":"[39]"},{"why":"Justifies the alignment between GNN message passing and BP, motivating the EW-GNN design.","marker":"[27]"},{"why":"Establishes the short-block-length performance gap and the URLLC context that motivate the design target.","marker":"[1]"}],"fun_headline_variants":["DRL-designed codes with GNN decoder outperform LDPC and BCH","AI-crafted codes beat classic error-correcting schemes","GNN decoder generalizes across code lengths without retraining","DRL and GNN codes gain up to 1.81 dB over classical schemes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reward used during code search—simulated BER at one SNR (6 dB) with a fixed 8-iteration BP decoder, or a hand-weighted combination of minimum distance and 4-cycle count—correctly ranks code quality at all operating SNRs and under the final decoder.","fun_headline_variants_meta":{"raw":{"variants":["DRL-designed codes with GNN decoder outperform LDPC and BCH","AI-crafted codes beat classic error-correcting schemes","GNN decoder generalizes across code lengths without retraining","DRL and GNN codes gain up to 1.81 dB over classical schemes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000848,"raw_usage":{"total_tokens":3690,"prompt_tokens":943,"completion_tokens":2747,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":2670}},"tokens_in":559,"tokens_out":2747,"duration_ms":218812,"temperature":1.0,"reasoning_tokens":2670,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:54:57.594463+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the minimum Hamming distance of the auto-encoder code $H^*_{AE}$ and evaluate its BER at SNRs well below and above 6 dB (e.g., 3 dB and 8 dB) with the EW-GNN decoder and with MLD; if the code's low-density, low-cycle structure makes it worse than the reference LDPC under MLD away from the training point—as already happens for the BER-reward-only code $H^*_d$—then the single-SNR reward is not a valid proxy and the headline gains are operating-point artifacts.","supporting_citations":[{"cited_title":"Short block length LDPC codes for TC synchronization and channel coding,","cited_arxiv_id":null,"evidence_quote":"Supplies the standard (32,16) LDPC code used as state initialization and as the main baseline for the auto-encoder comparisons."},{"cited_title":"Neural belief propagation auto-encoder for linear block code design,","cited_arxiv_id":null,"evidence_quote":"Provides NBP-AE, the learned auto-encoder baseline the proposed system must beat, and the GNBP decoder compared in the simulation figures."},{"cited_title":"Deep learning methods for improved decoding of linear codes,","cited_arxiv_id":null,"evidence_quote":"Defines the neural belief propagation decoder used as a benchmark for EW-GNN on BCH and LDPC codes."},{"cited_title":"Graph neural networks for channel decoding,","cited_arxiv_id":null,"evidence_quote":"Introduces the GNN decoder over Tanner graphs that EW-GNN extends with edge weights."},{"cited_title":"A scalable graph neural network decoder for short block codes,","cited_arxiv_id":null,"evidence_quote":"Prior work by the authors establishing the scalable GNN decoding structure whose parameter count does not change with code length."},{"cited_title":"Continuous control with deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Provides the deep deterministic policy gradient algorithm underlying the DRL code designer."},{"cited_title":"Neural message passing for quantum chemistry,","cited_arxiv_id":null,"evidence_quote":"Supplies the message-passing neural network formalism used by both the MPMNN actor-critic and the EW-GNN decoder."},{"cited_title":"What can neural networks reason about?","cited_arxiv_id":null,"evidence_quote":"Justifies the alignment between GNN message passing and BP, motivating the EW-GNN design."},{"cited_title":"Short block-length codes for ultra-reliable low latency communications,","cited_arxiv_id":null,"evidence_quote":"Establishes the short-block-length performance gap and the URLLC context that motivate the design target."}],"review_version":1}