acceptodds
Under review as a conference paper at ICLR 2027

Conditional Diffusion Decoders for Error-Correcting Codes

Abstract

We introduce a diffusion framework for neural decoding of error-correcting codes (ECCs). We interpret decoding as a conditional diffusion process that progressively predicts codeword bits while treating the physical channel observation as fixed side information. At each stage, a neural model estimates bitwise posteriors conditioned on the physical observation and the current noisy representation of the codeword. We focus particularly on masked diffusion, for which the reverse process naturally induces a list decoder inspired by the Maxwell construction of Measson et al. Rather than committing to a single uncertain decision, the decoder maintains multiple candidate trajectories, removes parity-infeasible candidates, and ranks the remaining candidates using accumulated reverse-transition log-probabilities. We evaluate the framework across multiple code families on binary phase-shift keying (BPSK) over additive white Gaussian noise (AWGN). Experiments show competitive decoding performance across Polar, low-density parity-check (LDPC), and Bose–Chaudhuri–Hocquenghem (BCH) codes. The list-augmented decoder consistently improves upon the corresponding single-trajectory diffusion decoder and, in most evaluated settings, improves upon available neural decoding results.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
error-correcting codes, neural decoding, conditional diffusion, list decoding, information theory

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_MRCfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a conditional masked diffusion framework for decoding error-correcting codes. It uses the observations received from the physical channel as fixed conditioning information and employs a CrossMPT to estimate bit-wise posterior probabilities. The authors further introduce a beam-search decoder inspired by the Maxwell construction. The decoder branches at low-confidence positions and prunes candidate paths using parity-check constraints. The experiments cover different code families (e.g., Polar, LDPC, and BCH codes), code rates, and SNRs. The results show that the Maxwell decoder outperforms its single-path diffusion counterpart and achieves competitive decoding performance in most settings.

Critical strengths
  • The paper presents a conditional masked-diffusion formulation for decoding linear codes. It progressively revealing masked codeword bits, offering an alternative to the continuous denoising approaches of DDECC and SB-ECC.

  • The Maxwell-inspired beam decoder combines model-guided branching, parity propagation, and minimum-distance selection from the final candidate list. This extension substantially improves upon the single-trajectory decoder and achieves lower reported BER than SB-ECC in most evaluated BPSK-AWGN settings.

Critical weaknesses
  • As currently presented, the sampled loss used in Eq. (11) does not appear to be an unbiased estimator of Eq. (7). There may be a gap between the theoretical analysis and the optimization objective. Correct me if I am wrong.

  • The base model, which uses masked diffusion, performs worse than the SOTA model SB-ECC and, in some cases, even the denoising diffusion model DDECC.

  • The weaker performance of the base model suggests that most of the performance gains may come from Maxwell beam search. However, the manuscript does not appear toprovide sufficient ablation studies or analysis of Maxwell beam search.

  • Since Maxwell beam search performs MAP-based selection within a restricted candidate set, more extensive comparisons with similar algorithms, including conventional non-NN methods, are needed. Examples include the original BP-based Maxwell decoder (whereas this work uses NN predictions in place of BP) and the family of ordered statistics decoding (OSD) methods.

Questions for authors

Please refer to the Weaknesses section.

4 comments
AuthorsOpenReview ↗

Response to Reviewer MRCf

W1. Whether the sampled loss in Eq. (11) is unbiased for Eq. (7)

We thank the reviewer for pointing this out. Submitted Eq. (11) omitted the masked-mean normalization 1/∣M∣1/|\mathcal M|, now corrected in Eq. (10). Although we did not explicitly claim unbiasedness, we agree that the connection requires clarification.

Appendix B shows that forcing a singleton when the mask is empty introduces an O(1/n)O(1/n) bias relative to the per-bit conservation-law. It derives the exact bias in Eq. (22) and an exactly unbiased mask-size reweighting in Eq. (23), retaining all masked labels. The reported training loss remains the unweighted version.

Changes to the manuscript:

  • Corrected Eq. (10) and linked Appendix B from Section 4.2.
  • Added Appendix B with the bias proof and exact reweighting correction.

W2. Standalone masked diffusion versus SB-ECC and DDECC

We agree that standalone masked diffusion generally performs worse than SB-ECC and, in some settings, DDECC. Most of our performance gain comes from beam search. Our conditional formulation separates physical-channel noise from auxiliary diffusion noise, allowing us to choose masking and construct a principled search over partial assignments. Our goal is to extend the mechanism of SCL decoding to a broader range of codes, accepting weaker single-path performance in exchange for better overall decoding performance.

The SC-to-SCL comparison illustrates this motivation. SC uses exact recursive bit-channel likelihood calculations rather than a learned model, yet without a list it performs worse than the neural methods in Figure 1. Retaining alternative decisions substantially improves performance toward the ML reference: exact probability calculations alone do not eliminate the limitations of committing to one path.

Single-path performance also depends on the reveal order. In unpublished follow-up work, optimizing the reveal schedule with reinforcement learning while keeping the released model weights fixed improved performance. For Polar$(128,64)$ at Eb/N0=4 dBE_b/N_0=4\,\mathrm{dB}, the negative log BER was 6.696.69 with random eight-bit reveals, 7.637.63 with confidence-based eight-bit reveals, and 9.19.1 with the optimized schedule, compared with 9.039.03 for SB-ECC. We share the unpublished RL results only to further emphasize the role of the decoding path. The comparison between random and confidence-based reveal orders already demonstrates this dependence using the same model weights and eight-bit reveal steps. Thus, weaker single-path performance does not by itself imply a poor posterior model.

As noted in our response to Reviewer 3Uou, we corrected the syndrome embedding to match CrossMPT (Park et al., 2025) and reran the experiments, including parity-based early exits. The revised results improve, but SB-ECC still has lower BER than our standalone decoder in 39 of the 42 comparisons in Table 1.

Changes to the manuscript:

  • Changed the title to “Beam Search for Error-Correcting Codes via Conditional Diffusion” to emphasize search.
  • Explicitly acknowledged the standalone decoder's weaker performance in Section 5.2.

W3. Insufficient Maxwell beam-search ablations and analysis

We agree that the contribution of Maxwell beam search and its components requires further analysis. We added a component ablation study in Appendix J (Table 8), comparing final codeword selection, parity propagation, and iterative completion on Polar(128,64) and CCSDS(128,64). Iterative completion improves performance on both codes. Minimum-distance selection improves performance substantially on Polar, with a smaller observed effect on CCSDS. Parity propagation has a negligible observed effect in these experiments.

We also added a comparison of BP with and without Maxwell beam search on CCSDS(128,64) and Polar(128,64) in Figure 5, with discussion in Section 5.2 and implementation details in Appendix D.3. This examines the contribution of search with a non-neural posterior estimator. Together, these comparisons clarify which components contribute to the gains and how the benefit of search depends on the posterior estimator and code structure.

Changes to the manuscript:

  • Added the Maxwell component ablation study in Appendix J (Table 8).
  • Added the BP versus BP-MBS comparison in Figure 5 and Appendix D.3.

W4. Comparisons with BP-based Maxwell decoding and OSD

We agree that comparisons with stronger conventional decoders are important. We added Appendix F (Tables 3 and 4), comparing CRC-free SCL with list size 64 for Polar codes and channel-LLR OSD of orders 1 and 2 for LDPC, BCH, MacKay, and CCSDS codes against the Maxwell results in Table 1 at Eb/N0=4 dBE_b/N_0=4\,\mathrm{dB}. The tables report the list sizes and OSD candidate counts explicitly: OSD-1 and OSD-2 enumerate 1+k1+k and 1+k+(k2)1+k+\binom{k}{2} candidates, respectively, whereas Maxwell and SCL use a maximum list size of 64.

We also added BP with and without Maxwell beam search on CCSDS$(128,64)$ and Polar$(128,64)$ in Figure 5. This replaces the neural posterior estimator with BP while retaining the search mechanism, directly examining the benefit of search with a non-neural estimator. On CCSDS, BP-Maxwell outperforms both neural diffusion baselines—plain conditional diffusion and SB-ECC—and approaches neural Maxwell. On Polar, BP provides poor posterior estimates; adding Maxwell beam search yields only a modest improvement and does not close the gap to the neural methods. This contrast shows that the benefit of search depends on the quality of the underlying posterior estimates. The results are discussed in Section 5.2, with implementation details in Appendix E.3. Additional Polar SCL comparison details appear in Appendix E.4.

Changes to the manuscript:

  • Added SCL-64 versus Maxwell for Polar codes in Appendix F, Table 3, and OSD-1/OSD-2 versus Maxwell for LDPC, BCH, MacKay, and CCSDS codes in Table 4, including decoder configurations and candidate counts.
  • Added the BP versus BP-Maxwell comparison in Figure 5, with discussion in Section 5.2 and implementation details in Appendix E.3.

Thank you very much for your continued engagement in the review process and for raising your useful concerns. Our responses are given below.

  • We believe that the changes to the paper, which were made in response to the reviewers' requests, are significant but do not make the new version substantially different. Although, the title was changed to highlight the paper’s main contributions, both the abstract and the introduction in the initial version stated that our goal was to build a beam search for ECC in a principled way, and the entire paper was built toward beam search from the beginning.
  • The changes to the method, such as early exits and completion attempts, were added in response to justified concerns about complexity. These changes do not change the character of the paper; they are refinements that improve the method we initially submitted.
  • Regarding more extensive experiments on beam search with other neural decoders, the reference methods use Gaussian noise and do not extend naturally to masking. One could devise a Chase-like decoder for those methods, but that would deviate from our method. We did compare BP with Maxwell and showed, surprisingly, that it beats all neural methods for the CCSDS code, but not for Polar codes. Comparing with other decoders that are not diffusion-based and predict bits would essentially amount to comparing the same method with a different backbone in place of CrossMPT. We are willing to add a comparison with the masked-diffusion-based decoder that the reviewer has in mind and include it in the paper.
  • Regarding more extensive experiments, we provide tail statistics in Appendix K.2. We are unsure what the reviewer means by alternative search strategies. We proposed a method and compared it with state-of-the-art methods, including the newly added comparisons with SCL and OSD and Appendix F. We believe this should address your concerns. We are also willing to add further ablations of training hyperparameters and decoder performance throughout training.
  • SB-ECC was trained for 3M steps with a batch size of 256. It is true that it is sensitive to the training budget. The original method was trained for 1.5M steps; we doubled this budget to reach saturation. We also allowed it to run for 100 diffusion steps, at which point we observed saturation in decoding performance. Based on our experiments and communication with the SB-ECC authors, training longer or using more diffusion steps will not yield the gains we have observed, and beam search is essential to achieving better results.
  • We added more list methods, such as SCL and OSD, which we present in Appendix F.

We appreciate the reviewer’s comments and found them very useful. But, we also believe that the principled design of beam search is important to the ICLR community, which has had many papers on neural ECC. Even though neural methods should, in principle, beat or match classical methods, they lag behind SCL for Polar codes. Our work has closed this gap and the way we did this is valuable to the community. We also want to emphasize that our empirical results are well above the significance threshold. We outperform the state of the art in all settings, especially for longer codes and whenever there is a mismatch (for Rayleigh, we improve by a factor of approximately 100).

I thank the authors for their rebuttal and clarification. I am now convinced that the manuscript has not undergone a substantial shift in research focus. I also recognize that conditional diffusion and MBS form an integrated framework.

However, I do not consider the experimental evaluation in the revised manuscript sufficient to fully address the concerns raised in my previous comment.


Rebuttal: “We are willing to add a comparison with the masked-diffusion-based decoder that the reviewer has in mind and include it in the paper.” The comparison I have in mind is not limited to another masked-diffusion-based decoder. Neural decoders that can output bitwise probabilities could be evaluated within the same MBS framework, analogous to the compared BP-MBS (by setting the chosen input bit by +1/-1). Such comparisons would help isolate the additional value of masked-diffusion conditioning from the gains due to search. I consider this particularly important because the proposed standalone decoder generally underperforms SB-ECC.


Regarding BP-MBS, the comparison in Figure 5 covers only two code configurations, which I consider insufficient to support broad conclusions. BP-MBS may remain competitive on other codes. In particular, on CCSDS, an LDPC code, BP-MBS already performs close to the proposed method. On Polar, its weaker performance is less surprising, given that BP is not a competitive decoder on Polar codes. Its weak performance on Polar therefore does not eliminate the need for broader comparisons involving other neural decoders combined with search or list decoding.

I recognize that neural decoders can remain valuable even when they underperform conventional decoders. I am not requiring the proposed method to outperform conventional decoders like BP-MBS. However, beyond the BP-MBS comparison, I would still like to see how standard neural decoders perform when combined with MBS.


Rebuttal “Regarding more extensive experiments, we provide tail statistics in Appendix K.2.” This remains my main concern. The appendix reports computational-work estimates and mean decoding times, but its beam-usage percentiles do not directly characterize actual tail latency. Furthermore, the end-to-end timing evaluation is limited to CCSDS(128,64), providing limited evidence about computational costs across code families.

Moreover, at 4 dB, Appendix K.2 explicitly reports that only 9 of the 1000 sampled frames activated MBS, which is consist with the reported number in Table 1 - Diffusion. The timing sample therefore contains very few search-triggering frames, potentially providing even fewer observations of rare, expensive search trajectories. This raises concerns about whether the sample adequately captures the search overhead and whether the reported mean costs are statistically stable. Larger samples, uncertainty estimates, and separate latency statistics for search-triggering frames would help assess these questions.

Finally, the runtime estimates in Appendix K.1 use RTX A5000 measurements, whereas the end-to-end timings in Appendix K.2 use an H200. Although the reason for using different devices is unclear to me, I respect the authors’ choice of experimental setup.


Rebuttal “Based on our experiments and communication with the SB-ECC authors”. I am not alleging a violation of the double-blind review policy. However, I consider this inappropriate.


Overall, my main concerns are:

  1. Whether masked diffusion is necessary for the reported performance gains.
  2. The need for a more comprehensive evaluation of MBS, particularly regarding computational complexity and decoding latency.

This work undoubtedly has the potential for publication at a top-tier conference such as ICLR. However, for the reasons discussed above, my overall assessment remains borderline, leaning toward weak reject. I leave the final decision to the Area Chair.

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

The paper formulates error-correcting code decoding as conditional masked diffusion, progressively predicting hidden codeword bits while keeping the noisy received signal fixed as side information. It introduces a Maxwell-inspired beam decoder that branches on uncertain bits, uses parity constraints to reject inconsistent candidates, and selects a final codeword from the surviving list. Experiments report improved error rates over several neural decoding baselines, with additional computational cost from maintaining multiple candidate paths.

Critical strengths
  1. This method attains higher -ln BER through the introduction of more computation; however, through parallel beaming, the increase in latency is controlled;
  2. The masked diffusion framework opens up the possibility of applying more techniques in neural decoding.
Critical weaknesses
  1. The computation incurred for Maxwell diffusion is heavy as a result of the beam search. To clarify the consequences of this, I wonder if it is possible to include operation counts (such as FLOPs) and batched latency (it is assumed that the current throughput is not practical in any implementation unless batched operations on GPUs are being considered) for the different methods.
  2. The source of improvement is not isolated. It is quite natural to assume this comes from the beam search; however, minimum-distance selection, parity propagation, or other factors may also contribute.
Questions for authors
  1. The comparison in Table 1 reports the BP decoder’s results for all code families, while conventionally BP is not suited for BCH and Polar codes. We do, however, notice that on line 418 the authors claim that the baseline methods include the SCL algorithm as well. For BCH codes, a reasonable candidate would be the OSD algorithm, as it is based on bit reliability and shares similarities with your masked diffusion to some degree (as suggested on line 142). You are also encouraged to provide classical decoders’ computational costs in Table 7.
  2. We notice that the paragraph at line 846 suggests that the current state and the observation are both being embedded in each layer of the crossMPT backbone. However, the provided source code suggests that these are only added to the initial input before the first layer, while later layers use ordinary layer norm. Please confirm this behavior.
  3. Equation 7 provides theoretical motivation, but its precise connections to the implemented training objectives and decoding algorithms are not established. Please explain what results or insights it should lead to.
  4. During training, the revealed bits are always correct, while in decoding, revealed bits are based on model predictions and could be wrong, with the positions based on confidence. This gap is not clearly quantified.
5 comments
AuthorsOpenReview ↗

Response to Reviewer kwGw

W1. Operation counts, batched latency, and throughput

We agree that beam search can incur substantial computational cost, and that operation counts and batched GPU measurements are needed to assess its practical throughput. To address this concern, we added FLOP counts, measured decoding times across batch sizes, and an analysis of beam usage. We also added parity-based early exits, as used in SB-ECC, to reduce unnecessary computation. Maxwell decoding now also attempts completion after the initial branching and after subsequent branching rounds. These completion attempts often find a valid codeword before the full beam budget is used, avoiding further search.

A single pass costs 0.520370.52037 GFLOPs for our model and 0.503430.50343 GFLOPs for SB-ECC. Counting all beam candidates on 1,000 matched CCSDS(128,64) frames gives 378.89378.89 versus 3.063.06 GFLOPs/frame at 2.5 dB, and 1.561.56 versus 1.981.98 at 4 dB. Thus, search is expensive at low SNR, but rarer beam activation reduces average predictor work below SB-ECC at 4 dB. Beam caps and step budgets are specified in Appendix K.2.

We also added measured GPU decoding times for independent-frame batch sizes of 1, 8, and 40 in Appendix K.2 (Table 11), comparing Maxwell, SB-ECC, and BP-50. For batches larger than one, the table reports batch latency divided by the number of frames, reflecting throughput.

Changes to the manuscript:

  • Section 4.3: parity-based early exits and intermediate completion attempts; Section 5.3 and Figure 6: their effect on neural evaluation counts.
  • Appendix K.1 and Table 9: beam-size comparisons and empirical beam-usage statistics.
  • Appendix K.2 and Table 10: single-pass and per-frame predictor FLOPs, with decoding budgets and accounting assumptions.
  • Appendix K.2 and Table 11: measured GPU decoding times for Maxwell, SB-ECC, and BP-50 at batch sizes 1, 8, and 40.

W2. Isolating gains from search, selection, and parity propagation

Most of the improvement comes from beam search. Final codeword selection also contributes: under AWGN, minimum-distance selection, which is the maximum-likelihood rule within the candidate list, performs better than selection based on the approximate model path score. The minimum-distance rule also performs well under Rayleigh fading, where it is not the exact maximum-likelihood rule. For example, on CCSDS(128,64) at 5 dB, Maxwell with minimum-distance selection reduces BER by more than 100× relative to both standalone diffusion and the published SB-ECC result (Appendix H, Table 6), exceeding two orders of magnitude. Parity propagation has a negligible effect on performance in our experiments. We inherited this component from the Maxwell decoder, but it does not play a substantial role in the gains achieved by our method.

Changes to the manuscript:

  • Added a component ablation study in Appendix J (Table 8), isolating final codeword selection, parity propagation, and iterative completion on Polar(128,64) and CCSDS(128,64), with the posterior estimator and received frames held fixed across variants.

Q1. Appropriate classical baselines and their computational costs

We thank the reviewer for highlighting the need for code-appropriate classical baselines and computational-cost comparisons. We added Appendix F (Tables 3 and 4), comparing SCL with list size 64 for Polar codes and channel-LLR OSD of orders 1 and 2 for BCH, LDPC, MacKay, and CCSDS codes against the Maxwell results in Table 1 at 4 dB. The tables specify the list sizes and OSD candidate counts. We also added a comparison of BP with and without Maxwell beam search on CCSDS(128,64) and Polar(128,64) in Figure 5, examining the benefit of search with a non-neural posterior estimator. Appendix K reports neural predictor FLOPs and measured decoding latency, including BP-50.

Changes to the manuscript:

  • Added the stronger-baseline comparisons in Appendix F: SCL-64 versus Maxwell for Polar codes in Table 3, and OSD-1/OSD-2 versus Maxwell for BCH, LDPC, MacKay, and CCSDS codes in Table 4.
  • Added the BP versus BP-Maxwell comparison in Figure 5, with discussion in Section 5.2 and implementation details in Appendix E.3. Additional Polar SCL comparison details appear in Appendix E.4.
  • Added predictor-work estimates in Appendix K (Table 10) and measured decoding latency for Maxwell, SB-ECC, and BP-50 (Table 11).

Q2. Conditioning in each CrossMPT layer versus only at the input

Thank you for pointing out this discrepancy. Your reading of the code is correct: conditioning on the current diffusion state and the physical observation is applied through an input adapter before the first CrossMPT layer, while the subsequent layers use ordinary layer normalization. The description of adaptive layer normalization at every layer referred to an earlier version of our model, which underperformed the current implementation. We inadvertently retained that description in the manuscript. We have corrected Appendix C in the revised version to match the implementation, including the direct contribution of the physical observation to the output logits.

Changes to the manuscript:

  • We revised Appendix C to replace the incorrect description of per-layer adaptive normalization with the implemented input conditioning adapter, ordinary layer normalization in the CrossMPT backbone, and direct observation-to-output connection.

Q3. Connection of Eq. (7) to training and decoding

Equation (7) connects bitwise prediction losses across masking levels to codeword-level conditional cross-entropy. It motivates our masked-bit training objective. We added Appendix B, which analyzes the bias of the sampled training loss relative to the conservation-law integral and provides an exact correction.

During decoding, these predictions determine reverse transitions and beam-search scores. With exact posteriors and sequential single-bit revelation, the accumulated log-probability equals the codeword log posterior by the chain rule. Appendix C analyzes the discrepancies from parallel revelation and approximate marginals. The path score is a discrete accumulation of local log probabilities, motivated by the connection between local prediction losses and global cross-entropy in Eq. (7).

Changes to the manuscript:

  • Added the training-loss bias analysis and exact correction in Appendix B.

  • Clarify the training connection and its assumptions in Section 4.2, and the path-score interpretation in Section 4.3, with references to Appendices B and C.

Q4. Training–decoding gap in revealed values and positions

We agree that the training–decoding gap deserves further examination. This problem, often termed exposure bias, is well known and widely studied in the diffusion-model community (see, e.g., Elucidating the Exposure Bias in Diffusion Models and Anti-Exposure Bias in Diffusion Models). Related concerns also apply to diffusion-based decoders such as SB-ECC: training evaluates the prediction loss on samples from the forward process, whereas decoding evaluates the learned field along model-generated trajectories.

Our beam-search mechanism helps address this problem by retaining alternative paths when predictions are uncertain. Confidence-based revelation likewise helps mitigate the problem rather than introducing it: it prioritizes bits whose values are nearly certain. In the limiting case of exact posterior certainty, revealing several bits in parallel introduces no error from ignoring their dependence, since deterministic marginals imply a deterministic joint distribution.

As a diagnostic, Appendix I (Table 7) reports the fraction of parity-valid outputs for LDPC(49,24) on the BSC with crossover probability 0.95. After adaptation, Maxwell produces valid codewords on 99.00% of the 3,616 evaluated frames, compared with 85.09% for diffusion alone. This measures output validity rather than decoding correctness or the frequency of invalid intermediate states; a fuller quantification of the training–decoding gap remains future work.

Changes to the manuscript:

  • Added the parity-valid output fractions for LDPC(49,24), including Maxwell, in Appendix I (Table 7).
Reviewer_3UouOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a conditional diffusion approach to error-correcting code decoding. The received signal is kept fixed while a masked diffusion model progressively reconstructs the codeword. The authors further introduce Maxwell beam search, which branches on uncertain bits and uses parity constraints to eliminate invalid candidates. Experiments cover several code families, including Polar, LDPC, and BCH. The beam search substantially improves decoding performance over the single-path decoder, achieving the best BER in 34 out of 42 comparisons. The paper also explores adaptation across different communication channels.

Critical strengths

The combination of conditional diffusion and code-specific constraints is well motivated. Maintaining multiple candidates for uncertain bits and pruning them using parity checks helps reduce the impact of early decoding errors.

The experimental results are convincing. The Maxwell decoder consistently improves over the single-path version across different code families, suggesting that the search strategy is not limited to a particular type of code.

Critical weaknesses

The main performance gains appear to come from beam search rather than diffusion itself. In Table 1, the single-path diffusion decoder underperforms SB-ECC in every setting. This makes the specific contribution of the conditional diffusion formulation less clear.

The inference cost is relatively high. Although the authors emphasize parallel processing of candidate paths, Table 7 shows that Maxwell decoding is still considerably slower than SB-ECC. This may limit its practical value for communication systems.

Questions for authors

Given that most of the performance improvement comes from Maxwell beam search, what does the diffusion formulation contribute beyond estimating bitwise posterior probabilities? Could the same search strategy be applied directly to other neural decoders?

Which decision rule is used for final codeword selection in the BSC(p=0.95) experiment? The minimum Euclidean-distance rule in Eq. (17) is derived for AWGN and does not correspond to maximum-likelihood decoding for the BSC(p=0.95) setting without output inversion. Clarifying this would help assess the channel adaptation results.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 3Uou

W1. Contribution of diffusion versus beam search

We agree. Most of the performance gain comes from beam search. In our experiments, SB-ECC generally performs better than standalone masked diffusion. Nevertheless, our conditional formulation sets the single path formulation, and is important for several reasons:

  1. Our goal is to imitate the mechanism of SCL across a broader range of codes. Thus, we ``sacrifice'' single path performance in order to obtain better overall performace.

  2. Consider SC decoding for polar codes. It uses exact recursive bit-channel likelihood calculations rather than a learned model. Yet, without a list, it performs worse than every neural method shown in Figure 1. Retaining a list of candidates substantially improves its performance toward the ML reference. Even exact probability calculations do not remove the need to retain alternative decisions: committing to a single path can itself be the bottleneck. This is precisely why we augment our conditional decoder with beam search.

  3. Weaker single-path performance also depends on the reverse noise schedule, i.e., the reveal order, and does not by itself imply a poor model. In unpublished follow-up work, we used the same model weights released with our diffusion model and optimized the reveal order. With these same weights, our single-path decoder surpassed SB-ECC, for which we did not optimize the schedule, as it is given by an ODE.

    For example, for Polar$(128,64)$ at Eb/N0=4 dBE_b/N_0 = 4\,\mathrm{dB}, SB-ECC achieves a negative log BER of 9.039.03. Our conditional diffusion model achieves 7.637.63 with a confidence-based schedule that reveals the eight most confident positions at each step, and 6.696.69 with a random schedule that reveals eight random positions at each step. The same model achieves 9.19.1 with an optimized schedule (we optimized it with RL). We share the unpublished RL results only to further emphasize the role of the decoding path. The comparison between random and confidence-based reveal orders already demonstrates this dependence using the same model weights and eight-bit reveal steps.

  4. SB-ECC is formulated around reversing Gaussian channel noise. Real-world channels need not be Gaussian. Our formulation decouples the physical channel from the diffusion process by treating the channel observation as a condition. This allowed us to choose masked diffusion, even though it performs worse than gaussian diffusion, to allow principled design of beam search.

We also clarify a correction to the implementation underlying the revised results. Our original implementation did not use the pure syndrome embedding employed by CrossMPT. After submission, we corrected the syndrome embedding to match the architecture described in Park et al., “CrossMPT: Cross-Attention Message-Passing Transformer for Error Correcting Codes,” ICLR 2025 and reran all experiments. These reruns yielded improved results for our decoder. In the updated Table 1, SB-ECC has lower BER than our standalone conditional masked-diffusion decoder in 39 of the 42 comparisons. This update does not change our main conclusion about the importance of beam search.

Changes to the manuscript:

  • We changed the title to “Beam Search for Error-Correcting Codes via Conditional Diffusion” to make the emphasis on search explicit.
  • In Section 5.2, we explicitly acknowledged the standalone decoder's weaker performance relative to SB-ECC.

W2. Inference cost and practical value

We agree with this observation. In our initial implementation, we did not allow early exits based on parity checks, as used in SB-ECC. Early exits were included in the rerun reported in Table 1 and are used in all newly reported results. After adding early exits, we observed computational benefits: our decoder terminates after a single neural evaluation much more frequently than SB-ECC (which rarely finishes after one iteration). We added Figure 6, which shows the distribution of the number of neural function evaluations (NFE) for our decoder and SB-ECC. The analysis in Section 5.3 and Appendix G.2 shows that, at the higher SNR of 4 dB on CCSDS(128,64), our decoder has a lower estimated mean inference time than SB-ECC.

Changes to the manuscript:

  • We added an explicit description of parity-based early exits and clarified that they are used in the Table 1 rerun and all newly reported results.
  • We added the NFE distributions for MBS and SB-ECC in Figure 6.
  • In Section 5.3, we discuss how early exits affect decoding cost at 2.5 and 4 dB.
  • In Appendix G.2 and Table 6, we report the corresponding estimated mean inference times.

Q1. What does diffusion contribute, and can other neural decoders use the search?

The conditional diffusion formulation provides the following:

  1. It decouples the channel noise from the diffusion noise.

  2. It guides the construction of the path metric. With exact posterior marginals and sequential single-bit revelation, the accumulated log-probability score equals the codeword log posterior. Appendix B shows that the expected negative path score upper-bounds posterior entropy, with additional penalties from parallel prediction and marginal approximation error.

  3. The conditional formulation through conservation laws led us to the Maxwell construction of Méasson et al.

More generally, any decoder that provides bitwise posterior estimates conditioned on the received observation and the candidate's partial assignments can be integrated into MBS. In Figure 5, we replace the neural model with BP using 50 iterations. For the CCSDS code, whose Tanner graph is locally tree-like, BP-MBS performs similarly to neural MBS and outperforms the standalone neural baselines. For the Polar code, however, BP provides poor posterior estimates. MBS offers only a modest improvement and does not close the gap to the neural methods. This comparison shows that MBS can be used with other posterior estimators, but its performance still depends on the quality of their estimates.

Changes to the manuscript:

  • In Section 2, we clarified how conditioning on partial assignments supports beam search over decoding trajectories.
  • We added the BP-MBS comparison in Figure 5 and discussed it in Section 5.2, showing that the same search can use a different posterior estimator and that its gains depend on the quality of the estimates.
  • In Appendix D.3, we describe the BP-MBS implementation, including conditioning on the candidate's revealed bits, replacing each neural-model call with 50 iterations of sum-product BP, and using the same search settings for both estimators.

Q2. Final codeword selection for BSC(p=0.95p=0.95)

Thank you for pointing out this ambiguity. In the BSC(p=0.95p=0.95) experiment, we used only the accumulated model log-probability score for final codeword selection. We selected the highest-scoring parity-valid candidate among the final beam and the valid completions saved during decoding. We did not use the minimum Euclidean-distance rule or invert the received observations. You are correct that the AWGN minimum-distance rule would not implement ML decoding. Our selection was based on the learned path score and should therefore be understood as approximate posterior-based selection, not exact BSC maximum-likelihood decoding.

Changes to the manuscript:

  • We moved the BSC example to Appendix G, “Adaptation to a Binary Symmetric Channel,” to make room for other observations in the main text.
  • In the appendix's “Final-selection rule” paragraph, we explicitly state that this experiment selects the parity-valid candidate with the maximum accumulated model log-probability score, including completion, and refer to the score-based decision rule in the main text.

Thank you for the detailed responses and additional experiments. My main concerns have been adequately addressed. I will maintain my rating of 3 (Weak Acceptance).

Thank you for your careful reading and for confirming that your main concerns have been adequately addressed. Your constructive feedback helped us clarify and improve the paper. We appreciate your constructive engagement and support.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision

We thank the reviewers for their constructive feedback which helped us to improve the paper. We revised the manuscript to clarify the role of conditional diffusion and beam search, strengthen the comparisons, and quantify computational cost. The main changes are:

  1. Clarified the contribution. We changed the title to “Beam Search for Error-Correcting Codes via Conditional Diffusion” and acknowledged that most of the performance gain comes from search, as we intended and expected.

  2. Updated the experiments. We identified a mismatch between our CrossMPT implementation and the architecture described in the literature. To address this mismatch, we added the syndrome embedding to our implementation, retrained the models, and reevaluated all codes. We also updated the search configuration to include intermediate completion attempts and a maximum beam size of 64. Appendix D describes the architecture, and Appendix E.2 documents the evaluation protocol. MBS achieves the best reported results among the methods compared in Table 1 across all 42 code–SNR settings. To the best of our knowledge, our method achieves state-of-the-art performance in these settings, including in comparison with methods submitted to ICLR 2027.

  3. Clarified the theory–implementation connection. Appendix B derives the training-loss bias and an exact reweighting correction. Appendix C analyzes the effects of parallel revelation and approximate marginals on the expected decoding path score.

  4. Added stronger baselines and component ablations. Appendix F (Tables 3–4) compares SCL-64 and channel-LLR OSD-1/OSD-2 with the Table 1 results at 4 dB. Our method is competitive with SCL-64 and OSD-2 in these comparisons. Figure 5 compares our method to BP with and without Maxwell search. Appendix J (Table 8) isolates final codeword selection, parity propagation, and iterative completion, while Appendix K.1 examines beam size and usage.

  5. Expanded the computational-cost analysis. Section 5.3 and Figure 6 examine parity-based early exits and neural evaluation counts. Appendix K.2 reports predictor FLOPs (Table 10) and measured GPU decoding times for MBS, SB-ECC, and BP-50 at batch sizes 1, 8, and 40 (Table 11), showing how cost depends on SNR and batching.

Detailed responses to each concern are provided in the individual reviewer replies.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.