acceptodds
Under review as a conference paper at ICLR 2027

I-FGMP: Factorized Gated Message Passing for Efficient LDPC Decoding

Abstract

Decoding low-density parity-check (LDPC) codes is a fundamental task in modern communication systems. Recent neural LDPC decoders have achieved strong decoding performance, but often require relatively large models and costly inference. In particular, diffusion- and score-based decoders can improve with additional inference steps, yet this test-time scaling incurs substantial computational cost. In this work, we propose Iterative Factorized Gated Message Passing (I-FGMP), which combines neural adaptability with two established principles of conventional message-passing decoding: sparse graph-local message exchange and reuse of a common update rule across iterations. Its core FGMP operator decomposes adaptive message exchange into source-side emission modulation, graph-structured aggregation, and destination-side acceptance gating. This design provides state-dependent message control without computationally expensive pairwise interaction modeling, while repeated application of the shared operator keeps the parameter count independent of inference depth. We train I-FGMP by finite-depth weight-tied unrolling, with supervision applied to the output state after each iteration. Experiments focused primarily on short codes show that I-FGMP trained for 12 iterations outperforms recent Transformer-based decoders with less than 1/10 the parameters and at most 4/5 the floating-point operations (FLOPs) of all baselines. When trained for 30 iterations, I-FGMP continues to improve with additional test-time iterations beyond its training depth, achieving performance comparable to the best diffusion- and score-based decoders while requiring less than 1/10 the parameters and roughly 1/3 the maximum allowable test-time FLOPs of the most competitive baseline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
neural LDPC decoding, error-correcting codes, message-passing neural networks, iterative inference

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_qps8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes I-FGMP, a neural BP-type decoder for LDPC codes. The main idea is to reuse common update rules across iterations. The authors show improvements over other neural-BP decoding baselines, mainly on six short LDPC codes. The approach is of some interest, but the contribution is very minor.

Critical strengths

The proposed approach improves on the selected baselines.

Critical weaknesses

The main weakness is that the contribution is too limited and lacks sufficient depth to warrant publication at ICLR. The shown improvements with respect to other neural BP decoders are promising, but the contribution is simply too narrow.

The construction is largely heuristic, with neither theoretical analysis nor novel insight.

For a paper that evaluates the performance of LDPC decoders, the results should be reported in terms of FER curves (alternatively BER curves, but FER are preferred) as a function of Eb/N0, as it is customary in any coding paper. The current way to present the results wouldn't be acceptable in more traditional venues for coding papers, and ML conferences should not be an exception. One needs to quantify the gains (in Eb/N0) and assess the behavior in the waterfall and error floor regions.

The main experiments concern lengths 49-128. But for these block lengths, most of the codes reported are not competitive. Polar codes (with CRC and list decoding), BCH codes with OSD, or other LDPC codes (potentially with OSD), would be better alternatives. As an example, only in the early times of LDPC codes one would report results on the MacKay code. MacKay codes are not competitive and hence they should not be considered when evaluating performance.

The comparison is insufficient given the computational cost of the proposed decoder. Comparisons with LDPC+OSD and other strong alternatives such as BCH+OSD and polar codes with CRC and list decoding, and a complexity discussion, are needed.

Questions for authors

Can the authors provide FER and BER curves, with gains measured in Eb/N0E_b/N_0 at fixed error rates?

2 comments

Thank you for your time and effort in reviewing our paper.

Response to W1, W2:

We have rewritten Section 4 and added Section 4.1.1 to clarify the architectural contribution and its motivation. Our core contribution is the FGMP operator, which implements adaptive message exchange through node-wise modulation and gating around fixed graph aggregation, avoiding explicit pairwise score evaluations.

Eq. (11) reveals that node-wise modulation and gating induce a separate rank-one reweighting of graph edges in each feature channel, providing interaction diversity across channels. This factorization is motivated by LDPC sparsity and local tree structure: Appendix B proves that, when the node factors vary freely within their activation ranges, arbitrary positive edge weights are representable if and only if the bipartite support graph is a forest. Higher-rank interactions therefore offer no additional edge-weighting flexibility within such isolated tree neighborhoods. However, cycles can make a single rank-one component insufficient. Channel mixing addresses this limitation by combining components across channels; when those channels carry the same message, their combination can form a higher-rank interaction, illustrating how message diversity can be traded for interaction flexibility. Because FGMP jointly learns messages and interactions, it can balance these two properties together with channel-wise interaction diversity from data, rather than fixing the allocation through attention-head partitions. The nonlinear modulator and gate further enhance expressiveness, while their elementwise implementation around fixed graph aggregation preserves computational efficiency without explicit pairwise score evaluations. More details are provided in Section 4.1.1.

The empirical results support this design: I-FGMP-12 outperforms the compared fixed-computation neural decoders at all 18 operating points in Table 1, using fewer than 10.0% of their parameters and at most 80.0% of their FLOPs. The expanded evaluation on five medium- and long-length codes further shows that I-FGMP-Scale generally outperforms BP-100, with substantial BER improvements at higher Eb/N0E_b/N_0 on most codes (Appendix F.2, Table 11). For example, on the longest code, LDPC$(1056,880)$, at 5.005.00 dB, its estimated BER is approximately 20.2% of BP-100's. Hardware measurements also demonstrate practical efficiency gains: on LDPC$(121,60)$ using an RTX 2080 Ti at B=1024B=1024, I-FGMP-12 achieves 1.91×1.91\times and 1.20×1.20\times the throughput of CrossMPT and EfficientMPT, respectively, while reducing peak allocated GPU memory by 79.0% and 36.7% (Appendix I, Table 13).

Response to W3, Q1:

We have added conventional BER and FER versus Eb/N0E_b/N_0 curves for all eleven codes in Appendix H (Figures 5–8). Table 12 reports Eb/N0E_b/N_0 gains over BP-100 at BER=10−4\mathrm{BER}=10^{-4} and FER=10−3\mathrm{FER}=10^{-3}, estimated by linear interpolation in log error rate between adjacent positive samples, without extrapolation. I-FGMP-Scale achieves positive gains on every code at both reference rates, with median gains of 0.2400.240 and 0.2100.210 dB, respectively. The curve trends suggest that larger gains may be obtained at lower target BER and FER.

Due to limited time and computational resources, we have been unable to fully characterize the behavior in the waterfall and error-floor regions through simulations. We will make effort to obtain the corresponding simulation results after the rebuttal period.

Response to W4, W5:

The six short codes, including the MacKay code, follow the benchmarks used in the original neural-decoder papers, enabling direct comparison of decoding architectures on the same codes. We have also extended the evaluation to five medium- and long-length codes in Appendix F.2 (Tables 10–11), where I-FGMP-Scale generally outperforms BP-100 and shows substantial BER improvements at higher Eb/N0E_b/N_0 on most codes.

We have also added comparisons with OSD of orders 1 and 2 on LDPC$(49,24),LDPC, LDPC(648,540),andLDPC, and LDPC(1056,880)$ in Appendix G (Figure 4). OSD-2 achieves the lowest BER and FER on the short code. On both longer codes, I-FGMP-12 and I-FGMP-Scale achieve lower BER throughout the tested range and lower FER from 3.503.50 dB onward.

Appendix G also discusses complexity. At fixed code rate, hidden dimension, and iteration budget, I-FGMP costs O(n2)O(n^2) per frame with dense aggregation, or O(n)O(n) with sparse aggregation on bounded-degree graphs. Both compared OSD implementations use O(n3)O(n^3) dense elimination per frame to select the information set, in addition to candidate-search costs of O(n2)O(n^2) for OSD-1 and O(n3)O(n^3) for OSD-2. I-FGMP therefore has more favorable asymptotic scaling with block length than these implementations.

Dear authors,

Thank you for your detailed responses and for taking the time to address the reviewer's comments.

Unfortunately, my main concern remains: the contribution of this work is simply too limited to warrant publication at ICLR.

Every year, a large number of papers make relatively minor contributions that are nevertheless valid and potentially useful to the research community. However, not all such contributions meet the bar for a top-tier venue like ICLR. I firmly believe that this paper falls into this category. I have carefully considered the authors' responses, but they have not changed my assessment of the paper's overall significance. I remain confident in my evaluation and stand by my original recommendation.

Given that this fundamental concern remains unaddressed, I do not believe that further discussion of the other comments would materially change my overall assessment. To be clear, I do not question that this work may be suitable for publication in a more less selective venue. However, I do not believe it meets the publication threshold for ICLR.

Best regards,

The reviewer.

Reviewer_D4viOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an iterative model with weight sharing for LDPC decoding. Compared with other AI-based decoders, it offers advantages in bit error rate, computational cost, parameter count, and test-time scalability.

Critical strengths
  1. Tables 1-2 show that, compared with other AI-based decoders, I-FGMP achieves a lower bit error rate with less computation and fewer parameters.
  2. Performance improvements are still when test-time iterations exceeds the training's.
Critical weaknesses
  1. Section 4 is hard to read and seems AI-generated. It only describes what the model is, but does not explain why. I cannot see the motivation or insights behind each module. If you rewrite Section 4 during the rebuttal, I will take my time to read it again.
  2. The fact that the parameter is independent of iterations offers limited novelty, as any model with weight sharing can achieve this. The importance of test-time scalability is questionable: would increasing the training depth be sufficient to achieve the same or better performance?
Questions for authors
  1. When designing the network architecture, how did you draw on the algebraic structure of LDPC, ideas from BP, and other AI concepts?
  2. In addition to computational cost and parameter count, could you report the average latency, including that of BP-100?
  3. In Figure 3, the curve for LDPC(121,60) at 6dB does not saturate as the iterations increases. Could you analyze why?
  4. Could you provide an ablation study on dd?
  5. Table 3 shows that the actual iteration is very less. I suggest reporting the results at 3dB to illustrate the performance when the iteration is more.
4 comments

Thank you for your time and effort in reviewing our paper.

Response to W1, Q1:

We have rewritten Section 4 and added Section 4.1.1 to clarify the design rationale. Starting from the gated aggregation in Eq. (8), we show in Eq. (11) that source- and destination-side factors induce an implicit rank-one reweighting of graph edges. These factors vary across feature channels, producing distinct effective aggregation matrices and thereby enabling interaction diversity. The structure of LDPC graphs further motivates this factorization. Appendix B proves that arbitrary positive edge weights can be represented on forest supports when the node factors can attain the required values. This result provides theoretical support for rank-one interactions in locally tree-like LDPC graphs. However, rank-one factorization imposes constraints on edge reweighting in the presence of cycles. To improve interaction flexibility, we employ channel mixing to combine features processed under different rank-one reweightings. By jointly learning message representations and interaction weights, FGMP can adapt message diversity, interaction flexibility, and interaction diversity to the data. The nonlinear modulator and gate further enhance expressiveness, while node-wise computation avoids explicit pairwise score evaluations. Further details are provided in Section 4.1.1.

Beyond the motivations discussed above, our architecture also draws conceptual inspiration from belief propagation (BP), particularly its emphasis on controlling information flow during graph-based message passing. BP uses extrinsic messages to avoid immediately feeding information back to its source and thereby prevent double-counting, whereas our model regulates message emission and acceptance through learned modulation and gating. Nevertheless, the connection is conceptual, and the underlying mechanisms differ substantially.

Response to W2:

We agree that weight tying is not novel. We use it to enable iterative updates and test-time scaling; our core contribution is the operator architecture designed for this iterative framework.

Regarding the question, simply increasing training depth does not necessarily outperform test-time scaling. We previously conducted two experiments: (i) comparing training depths of 30 and 50 iterations; and (ii) fine-tuning the depth-30 checkpoint using Truncated Backpropagation Through Time (TBTT) with 2 or 7 30-iteration segments (TBTT-2 and TBTT-7). Gradients were computed within each segment and arithmetically averaged across segments before each parameter update, exposing the operator to states over 60 and 210 iterations, respectively. Interestingly, the original depth-30 model achieved the best overall performance in both experiments.

One possible explanation is that iterative decoding produces cleaner, easier-to-decode states at later iterations. Deeper training may therefore overemphasize these easier states, diluting the learning signal from difficult cases and potentially impairing decoding performance.

As these experiments were conducted at an early stage of our research and the corresponding records are fragmented, we have not included their results in the revised manuscript.

Response to Q2:

We have added hardware inference measurements in Appendix I (Table 13) on LDPC$(121,60)$ at Eb/N0=5E_b/N_0=5 dB using an RTX 2080 Ti. We report the mean latency of one inference call decoding a batch of BB codewords. The B=1B=1 values are measured directly. For B=64B=64 and 1024, we obtain the mean batch latency as L(B)=B/R(B)L(B)=B/R(B), where R(B)R(B) is the measured throughput in codewords/s.

At B=1B=1, the mean latencies are 1.03, 1.07, 0.951, and 12.6 ms for I-FGMP-12, CrossMPT, EfficientMPT, and BP-100, respectively. At B=64B=64, the corresponding latencies are 5.76, 11.5, 7.29, and 14.4 ms. At B=1024B=1024, they increase to 85.5, 163, 103, and 45.0 ms, respectively.

At B=1B=1, the three neural decoders have similar latencies, while I-FGMP-12 reduces latency by 91.8% relative to BP-100. At B=64B=64, I-FGMP-12 reduces latency by 49.9%, 21.0%, and 60.1% relative to CrossMPT, EfficientMPT, and BP-100, respectively. At B=1024B=1024, the reductions relative to CrossMPT and EfficientMPT are 47.6% and 16.7%, while BP-100 has lower latency (45.0 ms versus 85.5 ms). Thus, I-FGMP-12's clearest latency advantages over the neural baselines occur under batched inference, with reductions of 16.7–49.9% at the two tested batch sizes.

These results suggest that, at B=1B=1, fixed execution overheads and limited GPU utilization may obscure the latency benefits of reduced arithmetic complexity. As batch size increases, these overheads can be amortized and GPU utilization may improve, allowing I-FGMP-12's lower FLOP count to yield more pronounced latency reductions relative to the neural baselines. Compared with BP-100, I-FGMP-12 achieves substantially lower latency at B=1B=1 and B=64B=64, whereas BP-100 is faster at B=1024B=1024.

Response to Q3:

We expect performance to eventually saturate, as observed for other codes and SNRs. Fundamentally, for a fixed code and SNR, channel noise imposes a nonzero minimum decoding error probability, even under optimal decoding. The original Figure 3 may have shown continued improvement simply because the iteration range was not sufficiently large to reveal this saturation.

In the revised manuscript, we modified the training data distribution for I-FGMP-Scale to improve its overall performance and consequently retrained the model used in Figure 3. In the updated figure, the curve reaches the saturation region at a smaller number of iterations.

Response to Q4:

We have added a hidden-dimension sensitivity study for d∈64,128,256d\in\\{64,128,256\\} on three codes in Appendix D.2 (Table 6), with the remaining settings unchanged. Increasing dd improves −ln⁡(BER)-\ln(\mathrm{BER}) at all nine operating points. Relative to d=128d=128, d=256d=256 yields gains of 0.140.14–0.490.49, but requires approximately 3.53.5–3.7×3.7\times the parameters and 3.73.7–3.9×3.9\times the FLOPs. Conversely, d=64d=64 reduces model size and computation with some performance loss. Overall, d=128d=128 provides a practical balance between decoding performance and computational cost.

Response to Q5:

We have expanded the early-stopping statistics in Appendix C (Table 3) to cover 2 and 3 dB. At 3 dB, the mean iteration counts range from 4.48 to 16.40 under a 60-iteration cap and from 5.05 to 22.92 under a 100-iteration cap. At 4–6 dB, all six codes require fewer than five iterations on average. At 2 dB, the mean iteration counts increase, exhibit greater variability, and become more sensitive to the iteration cap. These results show that early stopping reduces computation for easier frames while allowing harder frames to use more iterations.

Thank you for response. I have raised my rating from 1 to 2.

However, the readability still concerns me. The newly added Section 4.1.1 is a theoretical explanation after design, not a motivation before and during design. A good writing should:

  1. Before design: formalize a baseline method in preliminaries. My understanding is that I-FGMP improves upon BP, so BP should be formalized.
  2. During design: Explain the purpose of each module. Why is it needed? What does it improve relative to the baseline? Section 4.1 is insufficient and only in the last paragraph.
  3. After design: provide a theoretical explanation. Section 4.1.1 has done so.

Appendix B is also difficult to follow. The graph-theoretic proofs should include schematic diagrams, showing the bipartite graphs, spanning trees, forests, cycles, etc, that you mentioned.

The entire paper needs a careful human review for readability, with reorganization, simpler wording, and illustrative examples where necessary.

We thank the reviewer for their detailed suggestions. We will carefully revise the manuscript to further improve its overall readability.

Reviewer_1VhHOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes I-FGMP, a lightweight iterative neural decoder for low-density parity-check (LDPC) codes, aiming to improve decoding accuracy while reducing model complexity and computational costs. The key idea is to combine the structural efficiency of traditional message-passing algorithms with the adaptability of neural networks. The proposed Factorized Gated Message Passing (FGMP) operator decomposes message exchange into source-side emission modulation, fixed graph-structured aggregation, and destination-side acceptance gating. By sharing the same operator across iterations, the model supports additional inference steps without increasing its parameter count. Experiments on short LDPC codes demonstrate superior decoding performance over several Transformer-based baselines with substantially fewer parameters and FLOPs.

Critical strengths
  1. This paper focus on trade-off between neural decoding accuracy and computational efficiency. The proposed factorized gating mechanism is a compelling way to incorporate structural priors and reduce the complexity of adaptive message passing.

  2. The weight-tied architecture enables additional decoding iterations without increasing model parameters. The observation that a model trained for 30 iterations can benefit from more iterations at inference time is particularly interesting, as it provides a flexible accuracy-computation tradeoff.

  3. The paper is generally well organized, and the method is clearly formulated. The ablation studies systematically examine each component of method. The results support the effectiveness of these design choices and help understand the contributions of different components.

Critical weaknesses
  1. Although the results on short codes are strong, the advantages become less consistent for medium or long length codes. This raises questions about the scalability and practical applicability of the approach to larger coding problems, where BP is already highly effective.

  2. While the reduction in parameters and FLOPs is impressive, the paper provides limited evidence regarding actual inference latency, memory consumption, and throughput on real hardware. Since computational efficiency is a central motivation, real hardware experiment would strengthen the evaluation.

  3. The combination of source and destination gating with fixed graph aggregation is interesting. However, gated message passing and recurrent graph networks have been extensively studied. This paper would benefit from a clearer explanation of which aspects of the proposed factorization are fundamentally different from existing GNN-based and neural BP decoders.

Questions for authors
  1. The experiment shows the less advantages on longer codes. Could the authors further explain why the performance gap relative to conventional BP decreases as code length increases?

  2. Could the authors provide actual hardware inference latency, throughput, or memory measurements? It would be helpful to understand whether the theoretical FLOPs improvements translate into practical computational savings.

  3. Could the authors provide additional information about training seed variability or uncertainty estimates for the reported BER results, particularly at high signal-to-noise ratios?

3 comments

Thank you for your time and effort in reviewing our paper.

Response to W1, Q1:

The original medium- and long-length evaluation included only I-FGMP-12. We have now added I-FGMP-Scale on all five codes in Appendix F.2 (Table 11). It generally outperforms BP-100, with substantial BER improvements at higher Eb/N0E_b/N_0 on most codes, suggesting that the earlier results of I-FGMP-12 were partly limited by the iteration budget.

Nevertheless, the Eb/N0E_b/N_0 gains of I-FGMP over BP-100, reported in Appendix H (Table 12), generally decrease with block length. As discussed there, this is consistent with less pronounced finite-length effects, where BP more closely approaches its asymptotic iterative-decoding performance and potentially leaves less room for improvement.

Response to W2, Q2:

We have added measurements on an RTX 2080 Ti in Appendix I (Table 13 and Figure 9). On LDPC$(121,60)$ at batch size 1024, I-FGMP-12 achieves 1.91×1.91\times and 1.20×1.20\times the throughput of CrossMPT and EfficientMPT, respectively, while reducing peak allocated memory by 79% and 37%. All three have single-codeword latencies of approximately 1 ms; the clearest advantages are therefore batched throughput and memory usage. I-FGMP-12 also has lower latency than BP-100, although BP achieves higher throughput at large batches and uses less memory.

For batch sizes B=64B=64 and 10241024, I-FGMP-12 achieves 1.27×1.27\times and 1.20×1.20\times the throughput of EfficientMPT, respectively, close to the FLOP-based speedup of 208/166≈1.25208/166\approx1.25. Against CrossMPT, the corresponding throughput ratios are 1.991.99 and 1.911.91, exceeding the FLOP-based prediction of 237/166≈1.43237/166\approx1.43. These measurements support the practical relevance of the FLOP reductions; differences from FLOP-based predictions may reflect memory traffic, kernel overhead, and omitted lower-order operations (Appendix I).

Response to W3:

We agree that gating and recurrent weight sharing are established ideas. Our central insight is that the sparse, locally tree-like structure of LDPC graphs may limit the benefits of costly pairwise interaction modeling. In FGMP, source-side modulation and destination-side gating induce an implicit rank-one reweighting of graph edges in each feature channel (Eq. (11)). Appendix B provides a structural justification: when the node factors vary freely within their activation ranges, a single rank-one interaction can realize every positive edge weighting if and only if the bipartite support graph is a forest. Therefore, higher-rank interactions offer no additional flexibility for positive edge weighting on isolated tree supports, motivating this rank-one reweighting for locally tree-like LDPC neighborhoods. However, cycles can make a single rank-one component insufficient. Channel mixing provides additional interaction flexibility by combining features processed through different rank-one reweightings. When multiple channels carry the same message, their combination can form a higher-rank interaction, illustrating how message diversity can be traded for interaction flexibility. More generally, jointly learning channel-wise messages and interactions allows FGMP to balance message diversity, interaction flexibility, and interaction diversity from data. The nonlinear modulator and gate further enhance expressiveness, while their node-wise implementation around fixed graph aggregation avoids explicit pairwise score evaluations. We have added Section 4.1.1, with supporting analysis in Appendix B, to explain this design rationale in detail.

This operator architecture is our core distinction from GNN-based decoders that explicitly maintain edge messages or features: FGMP instead maintains node states and applies node-wise modulation and gating around fixed graph aggregation. The distinction from model-based neural decoders is broader, since I-FGMP is model-free and learns state updates without retaining prescribed BP/MS update equations.

Response to Q3:

We have added a five-seed evaluation of I-FGMP-12 on LDPC$(121,60)$ in Appendix E (Table 7), reporting individual results and means with sample standard deviations for both BER and −ln⁡(BER)-\ln(\mathrm{BER}). The standard deviation of −ln⁡(BER)-\ln(\mathrm{BER}) is below 0.010.01 up to 4 dB and reaches 0.0930.093 at 6 dB; the relative standard deviation of BER rises from 0.830.83\\% at 4 dB to 9.39.3\\% at 6 dB. Overall, we consider this level of variability acceptable.

Thank you for the substantial revision.

The authors add longer-code results, hardware measurements, and BER analysis, while clearly stating where the method does and does not outperform BP. These clarifications make the scope of the claims more precise. Within that scope, the factorized gated operator and its test-time-scalable, weight-tied decoder make a meaningful contribution. I consider my concerns sufficiently addressed and raise my rating.

We thank the reviewer for their recognition of our contributions. We greatly appreciate their constructive feedback and support.

Reviewer_jTDvOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes I-FGMP, an efficient model-free neural LDPC decoder based on factorized gated message passing. By combining sparse graph-based message passing with parameter sharing across iterations, the proposed method reduces computational complexity while supporting test-time scaling. Experimental results demonstrate strong decoding performance on short LDPC codes with substantially fewer parameters and FLOPs than existing Transformer-based decoders.

Critical strengths
  • The paper presents an efficient model-free neural decoder that achieves a favorable balance between decoding performance and computational complexity.

  • The proposed method demonstrates strong decoding performance on short LDPC codes, outperforming existing Transformer-based decoders with significantly fewer parameters and FLOPs.

  • The iterative architecture with shared parameters enables test-time scaling without increasing model size, providing flexibility in the performance–complexity trade-off.

Critical weaknesses
  • The computational advantages of the proposed lightweight architecture could be particularly valuable for longer LDPC codes. However, the main experiments focus primarily on short codes, while Appendix E only compares the proposed method with conventional BP. Comparisons with existing neural decoders, such as CrossMPT and EfficientMPT, on medium- and long-length codes are needed to better demonstrate the practical benefits of the proposed approach.

  • Although the use of −ln(BER) is consistent with prior work, conventional BER curves would make the results easier to interpret and compare. Including BER plots, at least in the appendix, would improve the presentation.

  • Frame error rate (FER) is an important metric for evaluating practical decoding performance. Given that the proposed method supports iterative refinement and syndrome-based early stopping, its FER performance would be particularly interesting to investigate.

Questions for authors
  • Could the authors extend the medium- and long-length LDPC experiments in Appendix E to include CrossMPT and EfficientMPT, and report BER, FLOPs, and parameter counts? In particular, it would be helpful to understand whether the proposed method maintains its computational efficiency and decoding performance advantages as the code length increases.

  • Could the authors provide conventional BER versus (E_b/N_0) curves, at least in the appendix? While −ln(BER) is commonly used in related work, BER curves would allow a more intuitive assessment of the performance differences.

  • Could the authors report FER results and compare them with existing model-free neural decoders? In particular, does the iterative architecture with syndrome-based early stopping provide an advantage in FER performance?

3 comments

Thank you for your time and effort in reviewing our paper.

Response to W1, Q1:

Training and testing CrossMPT and EfficientMPT on medium- and long-length codes are computationally expensive. With limited computing resources and time, we could not complete the medium- and long-length simulations within the rebuttal period. We have obtained results only on LDPC$(49,24)$ and LDPC$(121,60)$, reported in Appendix H (Figure 5). We will add performance comparisons with both neural decoders on medium- and long-length codes in the final version.

Response to W2, Q2:

We have added conventional BER versus Eb/N0E_b/N_0 curves for all eleven codes in Appendix H (Figures 5–8). I-FGMP-Scale consistently improves on I-FGMP-12, which generally outperforms the BP baselines. At BER=10−4\mathrm{BER}=10^{-4}, I-FGMP-Scale achieves positive Eb/N0E_b/N_0 gains over BP-100 on every code, with a median gain of 0.240.24 dB (Table 12). The curve trends suggest that larger gains may be obtained at lower target BER.

Response to W3, Q3:

We have added FER curves for all eleven codes in Appendix H (Figures 5–8). I-FGMP-Scale consistently improves FER over I-FGMP-12, while BP-100 remains competitive with I-FGMP-12, particularly on longer codes. At FER=10−3\mathrm{FER}=10^{-3}, I-FGMP-Scale reduces the required Eb/N0E_b/N_0 relative to BP-100 on every code, with gains of 0.010.01–0.370.37 dB and a median of 0.210.21 dB (Table 12).

Satisfying the syndrome condition guarantees codeword validity, but not correct recovery. Whether syndrome-based early stopping is particularly beneficial for FER requires experimental verification. The BER and FER results in Appendix H suggest that it may not offer a greater benefit for FER reduction than for BER reduction.

Thank you for the rebuttal and additional experiments. I appreciate the inclusion of BER and FER curves, which address two of my concerns. However, the comparison with existing neural decoders on medium- and long-length LDPC codes, particularly in terms of decoding performance and FLOPs, remains missing. Since computational efficiency is a key contribution of this paper, I believe this comparison is important. I will maintain my original rating.

We thank the reviewer for their time and careful consideration. We are currently conducting the requested comparisons on medium- and long-length LDPC codes and will include the results in the revised manuscript.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Summary of Revisions and Core Contributions

We thank the reviewers for their thoughtful feedback. The main revisions are summarized below, with all locations referring to the revised manuscript and changes highlighted in red.

  1. Clearer design rationale and theoretical support (Section 4.1.1; Appendix B). We have clarified FGMP's design rationale and the sources of its expressiveness and efficiency, drawing on the local tree-like structure of sparse LDPC graphs. We also provide a supporting theoretical analysis focused on the expressiveness of rank-one edge reweighting.

  2. Stronger medium- and long-code results (Appendix F.2, Table 11). The added I-FGMP-Scale results show substantial BER improvements over BP-100, extending the performance advantages beyond short codes.

  3. Hardware efficiency (Appendix I, Table 13 and Figure 9). Compared with CrossMPT and EfficientMPT, I-FGMP-12 has similar single-codeword latency, while its FLOP reductions translate into higher batched throughput. It also uses substantially less peak allocated GPU memory.

  4. Additional evaluation and robustness evidence. We have added BER/FER versus Eb/N0E_b/N_0 curves and coding gains (Appendix H, Figures 5–8 and Table 12); hidden-dimension dd ablations (Appendix D.2, Table 6); and training-seed variability results (Appendix E, Table 7). We also provide OSD comparisons (Appendix G, Figure 4) and expanded early-stopping statistics (Appendix C, Table 3).

Core contribution:

Our core contribution lies in the novel and conceptually distinctive FGMP architecture, together with an iterative computational framework tailored to it. This combination achieves superior LDPC decoding performance compared with state-of-the-art neural decoders, while requiring substantially fewer parameters and less computation.

Dear authors,

Thank you for your active engagement during the rebuttal and the substantial revisions. I have a few remaining questions. I understand time is limited, so clarifications alone are fine if additional experiments are not feasible.

  1. Intermediate supervision. Per-iteration supervision is architecture-agnostic and could also be applied to CrossMPT. How would CrossMPT perform with the same supervision? Conversely, the ablation in Table 4 (final-iteration-only supervision) suggests that, without it, I-FGMP’s advantage over CrossMPT largely disappears. Would it be fair to view I-FGMP as comparable in performance to CrossMPT but substantially lighter? I would appreciate your view on how much of the gain comes from the FGMP operator itself.

  2. BP latency. The reported BP latency per iteration seems much higher than typical for a code of this length; even BP-12 is slower than I-FGMP-12. Could you describe the BP measurement setup in more detail for reproducibility (e.g., how CUDA Graphs were applied to the Sionna decoder, what is included in the timed region, software versions)? Since latency depends heavily on implementation, if you wish to claim an advantage over BP, a FLOP comparison under the same counting convention would be helpful.

Best regards, AC

Thank you for the time and effort devoted to evaluating our work.

Response to Q1:

We retrained CrossMPT from scratch on the three codes in Table 4 using intermediate supervision. Specifically, CrossMPT with intermediate supervision applies the original CrossMPT output head to the output of each of its six layers, with the same head shared across layers. The binary cross-entropy losses from these six outputs are uniformly averaged. The CrossMPT architecture remains unchanged. Training follows the same procedure as the original CrossMPT.

The table reports −ln(BER) (higher is better). Both the published and reproduced CrossMPT baselines use final-only supervision. The I-FGMP results are taken from Table 4.

CodeMethod4 dB5 dB6 dB
LDPC (49,24)CrossMPT (published)6.68‾\underline{6.68}9.52‾\underline{9.52}13.19
CrossMPT (reproduced)6.569.3613.30
CrossMPT (intermediate)6.619.3513.35‾\underline{13.35}
I-FGMP-12 (final only)6.519.2712.99
I-FGMP-12 (intermediate)6.859.8613.65
LDPC (121,60)CrossMPT (published)5.749.2614.78
CrossMPT (reproduced)5.729.2814.83
CrossMPT (intermediate)5.719.4214.57
I-FGMP-12 (final only)5.91‾\underline{5.91}9.67‾\underline{9.67}15.06‾\underline{15.06}
I-FGMP-12 (intermediate)6.029.8515.22
MacKay (96,48)CrossMPT (published)7.9711.7715.52
CrossMPT (intermediate)7.8411.4715.25
I-FGMP-12 (final only)8.39‾\underline{8.39}11.90‾\underline{11.90}15.63‾\underline{15.63}
I-FGMP-12 (intermediate)8.4811.9915.64

Overall, intermediate supervision does not consistently improve CrossMPT and can even degrade its performance. In contrast, I-FGMP-12 benefits consistently from intermediate supervision and achieves the best reported performance across all nine code–Eb/N0E_b/N_0 pairs. These results suggest that intermediate supervision may be better suited to I-FGMP than to CrossMPT. We offer two possible explanations for this difference.

First, I-FGMP and CrossMPT follow different computational paradigms. I-FGMP shares parameters across iterations and can, to some extent, be viewed as an iterative message-passing algorithm that repeatedly applies the same operator to refine its decoding state. In contrast, CrossMPT uses different parameters across layers, allowing them to learn distinct transformations and potentially develop progressively more abstract representations. The inductive bias introduced by intermediate supervision—that each iteration’s state should yield logits for the same binary multiplicative-noise variables through a shared readout—is naturally aligned with iterative refinement in I-FGMP. For CrossMPT, requiring every intermediate representation to support the same decoding objective may constrain its flexibility, particularly if different layers serve different roles in feature processing.

Second, I-FGMP may benefit more from intermediate supervision due to its greater depth. I-FGMP-12 uses 12 iterations, compared with six layers in CrossMPT. With more iterations, gradient information from the final output must propagate through a longer computational path. Intermediate supervision provides more direct gradient signals at different depths, potentially facilitating optimization of the shared parameters in I-FGMP.

Response to Q2:

We agree that latency comparisons are implementation-dependent. We have therefore rewritten the Measurement protocol paragraph in Appendix I of the revised manuscript to improve reproducibility, explicitly addressing the three aspects highlighted by the Area Chair: CUDA Graph implementation, the scope of latency measurements, and software versions.

To further investigate the latency difference between BP and I-FGMP, we collected additional GPU kernel statistics. As the GPU used for the manuscript's benchmarks was unavailable, we repeated the measurements on another physical RTX 2080 Ti of the same model. This also explains the slight differences between the latency results below and those reported in the manuscript.

Latency and kernel statistics.

DecoderLatency (ms)Kernels/requestTraced kernel median (μs)Kernels below 5 μs
BP-50.6952981.664100%
BP-121.5267041.664100%
BP-506.12729081.69699.99%
BP-10013.34858082.04899.95%
I-FGMP-121.0513342.17667.60%
I-FGMP (T=60T=60)4.87215822.27265.80%

At B=1B=1, BP-12 requires 1.526 ms per request, compared with 1.051 ms for I-FGMP-12, with non-overlapping 95% confidence intervals of [1.520, 1.537] and [1.036, 1.070] ms, respectively. The kernel statistics help explain this difference. BP-12 launches 704 kernels per request, more than twice the 334 kernels launched by I-FGMP-12. Moreover, all BP-12 kernels execute in under 5 μs, with a median of 1.664 μs, whereas only 67.60% of I-FGMP-12 kernels fall below this threshold, with a median of 2.176 μs. This indicates that BP distributes its computation across substantially more fine-grained kernels, primarily involving indexing, reductions, and elementwise operations. In contrast, I-FGMP concentrates more computation in matrix operations, with GEMM/GEMV accounting for 64.8% of its traced kernel time at T=12T=12. Together, these observations suggest that BP's more fragmented GPU execution contributes to its higher single-codeword latency despite its much lower arithmetic complexity.

Analytical operation counts per iteration. To complement the latency measurements, we analytically count the arithmetic operations and special-function evaluations required by each decoder per iteration, using a consistent counting convention. For both methods, we use n=121n=121, k=60k=60, and m=n−k=61m=n-k=61 as the analytical number of check nodes. Each iteration includes both update directions but excludes non-iterative input, embedding, and readout operations. Addition, subtraction, multiplication, and division each count as one operation, while a fused multiply-add (FMA) counts as two. A length-KK dot product counts as 2K−12K-1 operations (KK multiplications and K−1K-1 additions). Special-function evaluations are counted separately. The reference LayerNorm expression requires 7d+17d+1 arithmetic operations and one reciprocal square root (rsqrt) per row.

MethodOrdinary FLOPs/iterationSpecial-function evaluations/iteration
BP14,1892,684 exp + 2,684 log
I-FGMP, d=128d=12827,797,43023,296 sigmoid + 23,296 softplus + 182 rsqrt

These results demonstrate that lower measured latency does not necessarily imply lower arithmetic complexity. Accordingly, in the revised Appendix I, we have restricted our claim to single-codeword latency under the measured implementation and added this analytical operation-count comparison to clarify the distinction between measured latency and arithmetic complexity.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.