acceptodds
Under review as a conference paper at ICLR 2027

Recurrence Strikes Back: Efficient Neural Decoding Beyond Transformers

Abstract

Recent work has made attention-based neural decoders substantially more efficient, yet it leaves a more fundamental question open: is attention the right computational primitive for error correction at all? Error correction is naturally expressed as repeated constraint resolution over a sparse code graph, suggesting recurrence rather than token interaction. We introduce Neural Syndrome Flow (NSF), a recurrent neural decoder that replaces attention with sparse message passing over the Tanner graph. NSF maintains persistent variable- and check-node states and repeatedly applies a shared, state-conditioned transition across decoding iterations. Lightweight edge gating, syndrome feedback, and adaptive solver control further refine this recurrent process. We evaluate NSF across a range of LDPC codes, including MacKay constructions, with particular emphasis on long blocks up to length 2304. NSF outperforms EfficientMPT, the previous state-of-the-art attention-based neural decoder for long LDPC codes; at , it reduces BER by on the code and by on the code. Across the long-code suite, NSF uses - fewer FLOPs and - lower peak inference memory than EfficientMPT even at the full step budget, while syndrome-based early stopping at increases the FLOP reduction to -. A length-invariant shared-initialization variant uses only 152k trainable parameters, fewer than EfficientMPT, while retaining most of the decoding gain. These results show that structured recurrence can simultaneously improve decoding accuracy and computational efficiency for long-block neural decoding.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Channel coding, Error-correcting codes, Neural decoding, LDPC codes, Recurrent neural decoder, Message passing, Tanner graphs, Iterative decoding, Efficient inference

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rR59OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a variant of neural BP termed Neural Syndrome Flow (NSF), formulated as a learned recurrent solver. The authors show improvements with respect to BP as well as other neural BP decoders, although the gains are very limited (as shown in Figures 3 and 4).

Critical strengths

The proposed decoder shows some improvement (albeit marginal in some cases, and in some others it performs worse) than BP and other neural BP decoders.

Critical weaknesses

The main weakness is that the contribution is too limited and lacks sufficient depth to warrant publication at ICLR. The shown improvements are of some value, although the improvement in terms of BER is very limited in most cases (and the proposed approach performs worse than other neural BP for some BCH and polar codes). In any case, the contribution is simply too narrow.

In my view, a simple idea has been expanded to a full ICLR paper. I find very little in the proposed approach that is surprising or conceptually substantial.

Note that, improving upon existing methods does not, by itself, warrant a publication, particularly at a venue such as ICLR. Indeed, a lot of purely engineering work can improve upon the state-of-the-art, without constituting a significant research contribution.

The construction is largely heuristic, with neither theoretical analysis nor novel insight. The paper does not demonstrate a substantial advance in scientific understanding or methodology.

For a paper that evaluates the performance of LDPC decoders, performance results should be reported in terms of FER curves (alternatively BER curves, but FER is preferred) as a function of Eb/N0, as it is customary in any coding paper. So please, replace the tables that give the -ln(BER) as a function of SNR with FER curves. These provide a much clearer picture of decoding performance and make comparisons easier to assess.

Even if this tabular format has been used in previous neural BP papers, that is not a good reason to retain a poor way of presenting decoding results. These kind of results would not be appreciated in any serious coding paper. Authors should follow established practice in coding rather than perpetuate an inadequate presentation format.

For polar and BCH codes the comparison should include the state-of-the-art decoders, i.e., CRC+list decoding and OSD, respectively.

Questions for authors

See weaknesses.

Reviewer_Rmu7OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

A neural belief propagation framework is proposed that achieves a lower bit error rate with efficient computation. Although the actual computational cost of the implementation remains to be investigated, its use of a GRU for iterative decoding and its gating mechanism are novel ideas.

Critical strengths
  1. The bit error rate is lower than that of all previous methods for a series of LDPC codes, with a lower computational cost.
  2. The introduction of a GRU for iterative decoding, as well as the gating mechanisms and the use of error logits, is rather novel and might be helpful for future advances.
Critical weaknesses
  1. The presentation is rather unclear throughout Section 3.2. A reader might be unable to comprehend the computation flow from the text and Figure 2, which contains too many connections. Details are given in the first two questions.
  2. The practical efficiency claims are insufficiently validated: the FLOP accounting differs from the released implementation, and per-sample early stopping does not eliminate computation for completed samples within a batch. Measured latency or throughput is needed to substantiate the claimed savings.
Questions for authors
  1. The notation and illustration in Section 3.2 need further adjustment and simplification. For example, the parameters λ\lambda and μ\mu in Equation 4 are explained on line 341, but we are unable to determine from the paper how μ\mu takes effect inside the recurrent units RcR_c and RvR_v.

  2. Can we assume that the overall computation follows a simple message-passing decoding process, where messages are accumulated using gates and learned projections, with a bit error prediction vector zz providing extra information throughout the iterations?

  3. This method exhibits a significant reduction in computational resource consumption for neural decoding. Could the authors reconcile the reported FLOP savings with the released implementation, which uses padded neighborhoods, a dense floating-point syndrome check, and continues processing completed samples until the entire batch finishes? Please report measured latency or throughput to verify that the theoretical and early-stopping savings translate into practical efficiency. Could you also compare against a BP baseline, such as NVIDIA’s Sionna GPU BP decoder implementation?

  4. The comparison with EfficientMPT raises several questions. Table 2 shows inconsistencies with Table 1 in the EfficientMPT paper; a little more explanation might clarify these. (i) The parameter count differs across codes, while in the original paper it is constant at 1.09M, as also suggested by your Table 1. (ii) The FLOP counts are different. For example, LDPC(121,70) is reported as 0.406G, while the original paper reports 0.20G. This might be explained by the original paper counting one multiplication and one addition as one operation. However, for WiMAX LDPC(576,432), the original paper reports 0.92G, while Table 2 reports 3.06G. As you stated, this is because more layers are used for EfficientMPT. Is there any justification for this?

  5. Also, is it fair to compare EfficientMPT (feature dimension 128) with NSF (feature dimension 64)?

  6. Could you provide BER results for long LDPC codes?

  7. This is not a question, and the authors may leave it unanswered. I do hope the authors try to simplify this model in their future work. The current ablations cover only one LDPC code; this is a bit limited for drawing persuasive conclusions.

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

This paper studies whether compact recurrent computation on the Tanner graph can improve the accuracy-efficiency trade-off of neural error correction decoding. The proposed Neural Syndrome Flow (NSF) maintains variable- and check-node states, exchanges messages through state-dependent scalar edge gates, and feeds the soft syndrome residual of the current error prediction back into the recurrent updates. The same transition is shared across decoding iterations, with syndrome-based early stopping. A shared-initialization variant further removes the dependence of the parameter count on code length. Experiments show improved BER and lower FLOPs and inference memory than the evaluated EfficientMPT configurations, particularly on long LDPC codes.

Critical strengths
  • The paper addresses a practical problem. Exploiting the sparse code graph to reduce the cost of neural decoding is well motivated.
  • The long-code results are promising. The reported BER reductions on the two longest LDPC codes are substantial, while NSF uses considerably fewer parameters than EfficientMPT.
Critical weaknesses

The novelty relative to existing recurrent and graph-based neural decoders is not sufficiently established. Cammerer et al. (2022) already learn shared node and message updates on the Tanner graph, and their messages are recomputed from the current endpoint states. Therefore, keeping persistent state at the nodes is not by itself a clear distinction from this prior GNN. Tian et al. (2023) also learn reliability-dependent scalar edge weights, while Qian et al. (2026) combine graph attention with GRU updates. The paper discusses these works but does not compare against them experimentally. To my understanding, NSF’s contribution is mainly the specific combination of inexpensive message computation, dynamic syndrome feedback, and adaptive control. A matched comparison with these closer methods is needed to establish the gain from this combination.

Questions for authors

See weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.