acceptodds
Under review as a conference paper at ICLR 2027

CR-CCM: Coset-Reduced Concatenated-Code Modeling for Neural Decoding

Abstract

Transformer-based neural decoders have achieved strong empirical performance under a unified formulation: binary linear codes are represented by parity-check matrices, and decoding is treated as bit-level noise prediction. This unified formulation is convenient, but it treats different code families in much the same way and can hide their structural advantages. For concatenated codes, the full-binary representation likewise obscures the distinct roles of the inner and outer codes. We propose Coset-Reduced Concatenated-Code Modeling (CR-CCM), a neural decoding framework tailored to algebraic concatenated codes. CR-CCM reformulates full-binary neural decoding as residual prediction on the binary-expanded outer graph. A coset-coordinate reduction removes repeated inner-code constraints from the neural task. A local soft interface provides tentative outer symbols and bitwise reliability, and the neural decoder predicts the remaining errors in these symbols. Relative to EfficientMPT, CR-CCM-EfficientMPT generally maintains or improves BER while substantially reducing computational cost. Average and maximum reductions are 81.0% and 94.9% in FLOPs, and 49.8% and 73.8% in peak training memory. At , BER is reduced by on average and up to approximately .

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Neural decoding, concatenated codes, error-correcting codes, Transformer-based decoding, coset reduction, message passing, efficient decoding

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_GtXVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work focuses on decoding concatenated codes, which combine inner and outer codes. In the main experiments, the authors use brute-force MAP/ML search to obtain an optimal inner-code estimate and its bitwise reliability (e.g., for BCH(15,5)). Since the inner codes considered have small dimensions, this setting is reasonable. The resulting reliability and outer-code syndrome are then fed into an EfficientMPT decoder to correct residual errors in the outer code, such as an RS code.

Critical strengths
  1. Compared with EfficientMPT using a unified PCM for both inner and outer codes, the proposed method substantially reduces FLOPs while generally maintaining or improving BER performance.

  2. The experiments cover various combinations of inner and outer codes.

Critical weaknesses
  1. Since concatenated codes naturally support separate inner- and outer-code decoding, I would say the novelty of this work is incremental. The method combines brute-force inner-code MAP/ML decoding with an existing EfficientMPT decoder for the outer code.

  2. The experiments show substantial FLOPs savings over EfficientMPT, while the BER gains are modest in most cases. Moreover, the practical value of these computational savings may be limited, since full-binary EfficientMPT is not a mainstream approach to decoding concatenated codes.

  3. The “Classical” baseline in Table 1 uses family-matched inner decoders, while the proposed method uses optimal MAP/ML search for inner-code decoding. This comparison makes it difficult to determine how much of the performance gain comes from the inner decoder and how much comes from the neural outer decoder.

Minor

  1. I would not say that existing Transformer-based neural decoders “treat different code families in much the same way and can hide their structural advantages” as stated in the Abstract. Code-specific constraints are represented without loss in the parity-check matrix and reflected in the Transformer mask design. When viewed as an unrolled computation graph, the Transformer has a code-specific structure, in my opinion.
Questions for authors

Please refer to the weakness.

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

The paper studies neural decoding of algebraic concatenated codes, which have an outer code over F2kin\mathbb F_{2^{k_{in}}} (RS or NB-LDPC) and a binary inner code. Instead of running a generic Transformer decoder on the full binary parity-check matrix, it proposes CR-CCM, which works in three steps. First, a local soft interface (max-log, Chase or trellis) maps each inner block to a tentative outer symbol and a bitwise reliability vector. Then a syndrome-based neural decoder, working on the binary-expanded outer PCM, predicts the remaining bit errors. Finally, the corrected symbols are re-encoded by the inner code. A coset-coordinate decomposition (Prop. 1–2) shows that, with complete blockwise costs, this reduction keeps the feasible corrections and the weighted objective. Thm. 1 shows that every reconstructed block is a valid inner codeword. Across 32 concatenated-code configurations, the paper compares CR-CCM-EfficientMPT with full-binary EfficientMPT; NBP and classical references are reported on subsets of these configurations. Relative to EfficientMPT, it reports generally similar or better BER at 4–6 dB and average reductions of 75.8% in FLOPs and 42.2% in peak training memory. The framework is also applied to NBP and HyperBP.

Critical strengths
  1. Large, consistent cost reduction from a correct and general formulation. The reduction is exact under complete blockwise costs (Prop. 2), and every reconstructed block is a valid inner codeword (Thm. 1). Across 32 configurations, FLOPs and peak training memory drop by 75.8% and 42.2% on average compared with EfficientMPT (Tables 2, 4, 6). For a fixed outer code (which also fixes kink_{in}), the size of the reduced neural graph does not depend on the inner code length ninn_{in}. On RS(63,55;6)+RM(32,6), FLOPs fall from 19.32 to 0.98 G/codeword, and training time from 485.5 to 49.8 s per epoch. The same interface also reduces cost for NBP and HyperBP (Table 8). This matters because cost is a main obstacle for Transformer decoders on longer codes.
  2. The learned correction gives large gains where it was tested directly. In Table 3, with the same local interface, adding the neural decoder lowers BER at 6 dB from 4.56e-4 to 6.27e-6 on RS(31,25;5)+BCH(15,5) (about 73×). On NB-LDPC(107,99;8)+CONV(16,8) it goes from 1.05e-4 to 9.70e-7 (about 108×). BLER improves by about 104× and 184×. The reliability ablation shows that the bitwise reliability input is needed for these gains.
Critical weaknesses
  1. No comparison under matched soft information. The classical reference uses different inner-decoding procedures: Berlekamp–Massey and bounded-distance decoders for most inner codes (Sec. 4.1). The manuscript does not specify its reliability interface well enough to establish a comparison under matched soft information. For example, it does not say where the reliabilities for "reliability-aided errors-and-erasures" decoding come from, or what the NB-LDPC sum-product decoder receives. Therefore, the comparison with the classical baseline does not isolate the benefit of the learned outer decoder from differences in local processing and reliability information. This matters for practical significance. The classical reference is better than CR-CCM at 6 dB on three RS configurations: RS(63,55;6)+BCH(14,6), 10.62 vs 8.07 in −ln(BER); RS(63,55;6)+RM(32,6), 10.37 vs 9.54 (Table 1); and RS(34,28;8)+EXT-HAM(13,8), 11.04 vs 7.72 (Table 5 note). CR-CCM is better on the other nine tabulated cases. I would find this comparison more convincing if a classical soft outer decoder received the same local estimates on a few representative codes, with both BER and decoding cost reported. For RS codes, adaptive BP can use signed bit LLRs derived from the same local estimates. For the max-log interface, Lemma 2 gives Li,p=(1−2u~i,p)2σ2ρi,pL_{i,p} = (1-2\tilde u_{i,p})\frac{2}{\sigma^2}\rho_{i,p}; the conversion and scaling should be stated. For NB-LDPC codes, nonbinary sum-product normally uses symbol-level inputs. If these are complete symbol costs rather than probabilities built from the bitwise estimates, this extra information should be stated.
  2. The neural stage's own contribution is not measured on many configurations. Its benefit is clearly shown on the two configurations in Table 3. On several RS configurations, however, three different backbones give nearly identical BER. For example, on RS(63,55;6)+BCH(14,6), CR-CCM-NBP and CR-CCM-HyperBP give 5.09/6.41/8.07 (Table 7), and CR-CCM-EfficientMPT gives 5.10/6.41/8.07 (Table 1). Similar patterns appear for RS(107,99;8)+LDPC(16,8) and RS(63,47;6)+BCH(12,6). Local-only results are not reported for these codes. If Local-only matches CR-CCM within uncertainty there, a cheaper tested alternative exists for those configurations, and the cost savings relative to EfficientMPT would matter less there. For context, the BER gains are concentrated: the geometric mean over 32 configurations is 1.83×, the median is about 1.07×, and the two largest cases give 46% of the summed log improvement. This is not a flaw by itself, because maintaining BER at lower cost is part of the stated contribution. To better understand the neural stage’s contribution on these RS configurations, I would like to see Local-only BER and BLER with uncertainty estimates, along with counts of errors corrected and errors introduced by the neural decoder.
Questions for authors
  1. Please specify the classical reference and add a matched comparison.

    • (a) Where do the reliabilities and the erasure rule for RS decoding come from? What does the NB-LDPC sum-product decoder receive? Does Viterbi use soft metrics?
    • (b) Can you give the same local estimates to a classical soft outer decoder, as described in Weakness 1, for a few representative RS and NB-LDPC configurations, and report BER and decoding cost?
    • (c) Why is the classical curve for NB-LDPC(107,99;8)+CONV(16,8) almost flat between 5 and 6 dB (7.51 → 7.71)?
  2. Can you report Local-only BER and BLER (with uncertainty) and errors corrected vs. introduced for the RS configurations with tied backbone results (e.g., RS(63,55;6)+BCH(14,6), RS(107,99;8)+LDPC(16,8), RS(63,47;6)+BCH(12,6))?

  3. All results use one training seed (App. B). Can you report the variation over several seeds for a few representative configurations, both with small and with large gains?

  4. Please give the exact code constructions: the inner LDPC, convolutional, product and polar codes, the NB-LDPC parity-check matrices, and the field basis. These are needed for reproducibility.

  5. On RS(107,99;8)+POLAR(16,8), CR-CCM-EfficientMPT reaches 4.29 in −ln(BER) at 6 dB (Table 5). With the same local interface, CR-CCM-NBP and CR-CCM-HyperBP reach 7.85 (Table 7). Why does the Transformer backbone perform much worse on this code?

Suggestions:

  • Results above 6 dB would be informative. Between 5 and 6 dB, the classical reference improves faster than CR-CCM on RS(63,55;6)+BCH(14,6).
  • Report memory savings at the training batch size of 128, not only at batch size 1.
  • Test symbol-level inputs (complete costs DiD_i) instead of bitwise reliabilities, especially for NB-LDPC outer codes.
  • Add BER results for ECCT and CrossMPT, since they appear in the cost tables.
Reviewer_mmpHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CR-CCM, a neural decoding framework for algebraic concatenated codes (nonbinary outer code + binary inner code). Instead of running a Transformer decoder on the full binary parity-check matrix of the concatenated code, each inner block is first decoded locally to produce a tentative outer symbol and bitwise reliabilities. A syndrome-based neural decoder then predicts the residual errors on the binary-expanded outer graph, and the corrected symbols are re-encoded by the inner code. The paper shows that the coset-coordinate change of variables preserves the feasible set and weighted cost, and that reconstruction always yields valid inner codewords. Experiments on RS and NB-LDPC outer codes with various inner codes report average FLOPs and peak-memory reductions relative to full-binary EfficientMPT, and show mostly similar or better BER.

Critical strengths
  1. There is clear practical motivation and substantial efficiency gains. Applying generic Transformer decoders to full concatenated parity-check matrices creates massive graphs with redundant local constraints. Factoring out the inner code enables significant computational and memory savings, making neural decoding for long concatenated codes much more viable.

  2. The paper offers a flexible and well thought out framework. The soft interface cleanly connects different local estimation rules with multiple neural backbones, demonstrating applicability across various code combinations.

Critical weaknesses
  1. The paper has limited conceptual novelty. The core idea of decoding inner blocks locally to produce soft reliability information for an outer decoder is the foundational principle of classical concatenated decoding (dating back to Forney in the 1960s). Replacing the outer decoder with a Transformer is an incremental engineering combination of existing classical decoding pipelines with recent neural syndrome decoders.

  2. While the inner re-encoding step guarantees that every individual block is a valid inner codeword, Theorem 1 shows that the full output is a valid concatenated codeword only if the neural network's predicted flips clear the outer parity checks. Because the model predicts bit flips by simply thresholding its output numbers independently at zero, there is no hard constraint forcing the final output to satisfy the outer code. Whenever the neural decoder makes an error, the reconstructed word is not a valid concatenated codeword. I am curious, how often does the final output fail the outer checks, and could a classical algebraic completion step be added to ensure the outer code is always satisfied?

  3. The paper is missing ablation against classical soft outer decoding. The paper primarily compares CR-CCM against full-binary EfficientMPT, which is disadvantaged by having to learn both inner and outer code constraints simultaneously from scratch. To truly demonstrate the necessity and value of the neural outer decoder, the authors should compare against a baseline that uses the exact same local soft interface followed by a standard, non-neural soft outer decoder (e.g., standard Belief Propagation on the outer graph using the computed reliabilities).

  4. The performance of the proposed scheme is mixed compared to classical baselines. At moderate-to-high SNR (6 dB), standard classical algebraic decoders still outperform CR-CCM on several Reed-Solomon configurations (e.g., RS(63, 55; 6) + BCH(14, 6) ). Furthermore, the claimed 1.83x average BER improvement seems to be skewed by a couple of outlier code combinations, whereas on many other codes the improvement over simpler baselines is negligible.

Questions for authors
  1. How does the proposed method compare against a baseline that uses the same local soft-decision interface followed by a standard, non-neural outer decoder (e.g., classical soft-decision decoding or Belief Propagation on the outer parity-check matrix)? This would help isolate whether the performance gains stem from the neural outer model or simply from providing structured local soft information.

  2. Why does classical decoding continue to achieve lower BER than CR-CCM at higher SNRs for several Reed-Solomon outer codes? Does the neural outer decoder exhibit an error floor or struggle in the low-error regime?

  3. Results Table 3: In Table 3 (RS–BCH), "CR-CCM-EfficientMPT w/o reliability" reproduces "Local only" exactly in BER and BLER. Does this mean the no-reliability model learned to predict no corrections?

  4. Results Table 5: Several Table 5 rows show no gain (e.g., RS(32,26;6)+CONV(12,6), RS(63,55;6)+POLAR(16,6), RS(34,28;8)+PRODUCT(15,8)), meaning CR-CCM achieves identical BER to the baseline. On some of them, NBP-50 is better than both models. So, when does the neural residual correction actually help?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.