acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Diffusion for Neural Channel Decoding: Aligned Training and Heterogeneous Inference

Abstract

Neural channel decoders offer a common approach to decoding different families of linear block codes, but improving error correction often comes at the cost of slower inference. Diffusion decoders make this trade-off particularly apparent: they rely on repeated model evaluations, while existing training objectives do not always match the dynamics used at inference. We revisit diffusion decoding by aligning the training target with probability-flow dynamics and mapping the channel noise level directly to diffusion time. This yields DEER, a decoder that supports a fixed inference schedule and improves decoding performance in both one-step and multi-step settings. To reduce the cost of each step, we develop FAWN, a lightweight decoder with computation that scales linearly with the input sequence length and the number of nonzero entries in the parity-check matrix. We further observe that the high-capacity decoder is most useful in the noisier stages of a decoding trajectory, whereas lightweight decoders become competitive as the noise decreases. This motivates heterogeneous inference, which assigns DEER and FAWN to different stages of the same trajectory. Across several code families, the resulting decoders improve the measured bit-error-rate–latency trade-off, with heterogeneous inference matching or exceeding the performance of the high-capacity decoder at lower latency.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Diffusion Models, Channel Decoding, Error-correcting Codes, Probability Flow ODE, Soft Decoding

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_nZZNOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces DEER and its lightweight counterpart, FAWN, by aligning the training target with probability-flow dynamics and explicitly mapping channel noise levels to diffusion time. The motivation stems from inconsistencies between discrete predictions, continuous diffusion updates, and syndrome-based time indexing in existing neural decoders. The proposed framework is intuitive and well motivated, and the experiments demonstrate favorable BER–latency trade-offs across several code families compared with the evaluated neural baselines.

Critical strengths

1.The paper identifies concrete limitations of existing diffusion-based neural decoders and uses them to motivate its design choices. This makes the work intuitive and strongly motivated.

2.The author's method, building correspondence between channel noise level and diffusion time, together with the derivation of the regression objective, directly addresses the identified training–inference mismatch, which is rational and rigorous.

3.The effort to reduce efficiency is appreciated.

Critical weaknesses
  1. There appears to be an inconsistency between the legend of Fig. 2(a) and the corresponding text.

  2. Please discuss the scope and limitations of the proposed framework more explicitly. The BER evaluation primarily concerns short codes over binary-input AWGN channels, while the long-code experiments mainly demonstrate runtime scaling. It would be helpful to discuss the challenges of extending the approach to other channel models or higher-order modulation, and to distinguish the demonstrated computational scalability from the evidence for error-correction performance at longer code lengths.

Questions for authors
  1. There appears to be an inconsistency between the legend of Fig. 2(a) and the corresponding text.

  2. Please discuss the scope and limitations of the proposed framework more explicitly. The BER evaluation primarily concerns short codes over binary-input AWGN channels, while the long-code experiments mainly demonstrate runtime scaling. It would be helpful to discuss the challenges of extending the approach to other channel models or higher-order modulation, and to distinguish the demonstrated computational scalability from the evidence for error-correction performance at longer code lengths.

1 comment

We thank the reviewer for the positive assessment and the suggestions on presentation and scope. The figure inconsistencies are addressed in our response to Reviewer TcU4, and we will correct Figures 1 and 2 in the revised manuscript.

Our formulation considers BPSK transmission over a binary-input AWGN channel. Most error-rate experiments concern short block lengths. We will state this scope explicitly and distinguish these results from the long-code runtime measurements. Extending the formulation to other channels requires adapting the channel model, interpolation, and corresponding training target; the present analysis does not directly establish such extensions.

Although the heterogeneous decoder offers a favorable accuracy–latency trade-off on the evaluated short codes, its DEER component does not share FAWN’s linear-scaling attention implementation. We have not completed a comparable long-code error-rate evaluation under our current hardware and training configuration. The long-code experiments therefore demonstrate runtime scaling rather than decoding accuracy at those lengths.

Evaluating FAWN on longer codes is an important direction for future work. Its lower per-evaluation cost may also permit larger-capacity variants under a comparable computational budget. Whether such variants can match or outperform DEER at longer block lengths remains to be established experimentally.

Higher-order modulation would require revisiting the symbol mapping, likelihood representation, and diffusion formulation. We appreciate the reviewer’s suggestion and will discuss these challenges as limitations and directions for future work. These extensions are beyond the scope of the current BPSK-AWGN evaluation.

Reviewer_TcU4OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper revisits diffusion-based neural channel decoding, aligning the training target with probability-flow dynamics and mapping channel SNR to diffusion time. It introduces two decoders: DEER, which utilizes a CrossMPT backbone, and FAWN, a lightweight decoder built on a modified EfficientMPT backbone with a sparse kernel that achieves O(2N−K+h)O(2N-K+h) complexity. The paper also proposes hybrid inference, employing DEER for high noise steps and FAWN for low noise steps. This approach is based on the observation that DEER performs better at high noise levels, while FAWN maintains good performance with lower latency at lower noise levels. Experiments are conducted on Polar, BCH, LDPC, CCSDS, and MacKay codes at 4–6 dB, reporting −ln⁡-\lnBER and latency results.

Critical strengths
  • Hybrid inference, using a heavier model in the high noise steps and a lighter model in the low noise steps along one trajectory, is a genuinely interesting and practically useful idea that balances accuracy and latency.
  • Reducing the EfficientMPT's ECC attention complexity by using a sparse kernel is a clever and well-motivated idea.
Critical weaknesses

Presentation and reporting inconsistencies

The paper contains several inconsistencies and under-explained figures that make it harder to check the claims.

  • Fig. 2 against the text and the adopted loss. The page-4 text states that the MSE-trained model "exhibits a lower-error trajectory" than its BCE-trained counterpart and that the BCE model can "increase the number of bit errors in intermediate iterations". In Fig. 2(a), however, the BCE series (blue) ends lower than MSE (red) in all three panels and decreases monotonically. As submitted, the figure and the text are contradicting each other. This may simply be a legend/series swap, or a swap in the analysis. This does not affect the reported decoding results, which stand whichever loss is stronger; but if the analysis is the swapped one, it would also question the choice of MSE as the main training loss.
  • Diffusion-time indexing. Fig. 1 labels the noisiest end t0t_0 and the cleanest end tTt_T (time increasing toward clean), while Eq. (5) sets η0=0\eta_0 = 0 and η1=1\eta_1 = 1, i.e. t=0t = 0 is clean (time increasing toward noisy); the initial inference time is also named tinitt_{\mathrm{init}} on p.5 but t0t_0 on p.6, with the schedule written t1,…,tN=0t_1, \dots, t_N = 0.
  • Figure captions/conditions. Fig. 1's caption is not very informative. Fig. 4 is badly explained: the experimental conditions and the meaning of filled vs hollow markers are not stated (the difference can be figured out, but it should be stated explicitly).
  • Results presentation. Results are split between tables and figures and are not always consistent, and −ln⁡-\lnBER is not easy to read on first sight. Fig. 3, for example, should include the classical decoders as well: they are in Table 2, and on BCH the classical decoder is the strongest entry while having lower complexity.

Incremental contribution

The backbones (CrossMPT, EfficientMPT) and the diffusion machinery (DDECC) are prior work; the new elements are the aligned training/inference formulation, the backbone fixes, and heterogeneous inference. The contribution is therefore incremental rather than a new decoder class.

Questions for authors
  • In Fig. 2(a), which series is BCE and which is MSE? How should the plotted trajectories be reconciled with the sentence that the MSE-trained model has the lower-error trajectory?
  • Could the diffusion-time convention be stated explicitly and aligned between Fig. 1 and Eq. (5) (and tinitt_{\mathrm{init}} vs t0t_0 vs t=0t=0, etc.)?
  • Since adaptability to SNR is one of the paper's claims, have you considered extending the evaluation beyond the current 4–6 dB range? That range is narrow relative to the claim.
  • Have you considered reporting BLER alongside BER? In a coded system what ultimately matters is whether a block is decoded correctly; otherwise it is discarded and retransmission is requested, so a block-level error rate would be more informative than a bit-level one.
  • For Fig. 4(c)/(d): what are the conditions and step counts, what do filled vs hollow markers mean, and are those latencies comparable to Table 3 (the codes differ)?
  • Could the classical decoders also be plotted in Fig. 3, so the BER figures show the same baselines as the tables? On BCH they are the strongest entries in Table 2
2 comments

Presentation Inconsistencies

We thank the reviewer for identifying these presentation issues. We address them as follows.

Figure 1: We thank the reviewer for pointing out the ambiguity. In Figure 1, the subscripts index decoding iterations rather than indicate increasing diffusion time. Specifically, t0=tinit>t1>⋯>tT=0t_0=t_\text{init} > t_1 > \cdots > t_T = 0: decoding starts from the noisy observation at t0t_0 and proceeds toward the clean endpoint at tT=0t_T=0. Thus, the iteration index increases while diffusion time decreases, consistently with η(0)=0\eta(0)=0 in Eq. 5. We will clarify this distinction in the figure caption and in section 4.3, use tinitt_\text{init} consistently for the initial diffusion time, and denote the number of inference steps by TT to distinguish it from code length NN.

Figure 2: An incorrect plot file was included in the submission. We will replace it with the intended plot and revise the accompanying discussion to match the corrected figure.

We will expand the caption of Figure 4 to specify the experimental conditions and marker conventions:

(a–b) Single-pass latency of FAWN and EfficientMPT as a function of the number of nonzero entries in the parity-check matrix HH and the decoding token-sequence length, using a batch size of 256. (c–d) Decoding performance versus latency at Eb/N0=4E_b/N_0=4 dB. Hollow markers denote multi-step neural decoding with 16 steps; filled markers denote single-step neural decoders or classical decoders. Neural-decoder latency measurements in panels (c–d) use a batch size of 2,048.

We will add the classical decoders reported in Table 2 to Figure 3. In particular, the figure will reflect the strong performance of OSD on the evaluated BCH codes.

Evaluation across SNRs and block error rates

Our claim concerns allocating computation according to the noise level, rather than extending the applicable SNR range beyond that of prior methods. We initially focused on 4–6 dB to compare with commonly reported results. To supplement this evaluation, we reproduce CrossMPT, EfficientMPT, and SB-ECC and compare BER over 3–7 dB on two representative codes. We also report BLER at 4–6 dB on four codes below. We follow the original training and evaluation recipes where possible; some reproduced values differ from the published results.

CodesSNRCrossMPTefficientMPTSBECCOur Hybrid
BCH 63,363dB3.793.704.124.36
BCH 63,364dB5.094.955.686.04
BCH 63,365dB6.976.767.858.45
BCH 63,366dB9.639.2410.9711.51
BCH 63,367dB13.2312.8014.9315.39
Polar 64,323dB5.284.805.685.87
Polar 64,324dB7.326.587.717.94
Polar 64,325dB9.979.1810.3410.48
Polar 64,326dB13.3612.3313.5913.94
Polar 64,327dB17.5716.3817.3117.97

Our performance gains are preserved at lower SNR regions in those evaluated cases.

We will include a complete BLER table alongside BER for all tested codes in the appendix. The preliminary table below compares our heterogeneous decoder with the three reproduced baselines on four representative codes. Entries are −ln⁡(BLER)-\ln(\mathrm{BLER}), where higher values indicate better complete-codeword recovery.

CodesSNRCrossMPTEfficientMPTSB-ECCOur Hybrid
BCH (63,36)4 dB2.562.383.253.66
BCH (63,36)5 dB4.324.055.295.93
BCH (63,36)6 dB6.806.348.338.84
Polar (64,32)4 dB4.643.625.635.88
Polar (64,32)5 dB7.176.048.268.41
Polar (64,32)6 dB10.529.0311.4911.74
CCSDS (128,64)4 dB4.654.626.256.55
CCSDS (128,64)5 dB8.848.8111.0211.31
CCSDS (128,64)6 dB13.6213.6716.6313.23
Polar (128,64)4 dB2.472.195.628.29
Polar (128,64)5 dB5.374.829.1611.16
Polar (128,64)6 dB9.518.7213.0213.12

Our method remains competitive in the harder case of low snr regions, with 33dB as the noise level at t=1t=1, which is the highest noise during the training of our models.

Contribution

Our contribution combines an analysis of the training–inference mismatch in diffusion decoding, an aligned training/inference formulation, and an efficient implementation. DEER applies the formulation to the CrossMPT backbone, while FAWN combines it with architectural and computational optimizations to EfficientMPT. Heterogeneous inference then allocates these decoders across denoising stages. The experiments demonstrate improved accuracy–latency trade-offs relative to the evaluated baselines. For sparse LDPC codes, FAWN’s parity-check aggregation exploits the nonzero entries of HH, supporting linear scaling with code length when the number of nonzero entries grows linearly.

Thank you for your time and consideration. As the rebuttal period will close in approximately 24 hours, we would be grateful if you could let us know whether our response has addressed your concerns. If appropriate, we would also appreciate it if you could update your assessment in light of our response. We would be happy to clarify any remaining points or answer further questions before the discussion period closes.

Reviewer_pc6dOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper discusses diffusion-based neural channel decoding and introduces DEER, a decoder trained with a revised objective and a noise-to-diffusion-time mapping. It also presents FAWN, a more computationally efficient decoder, and a heterogeneous inference scheme that combines DEER and FAWN across decoding stages. Experiments on multiple code families demonstrate competitive decoding performance and improved latency–performance trade-offs relative to the evaluated baselines.

Critical strengths

The paper addresses a practical challenge in neural channel decoding: improving performance while managing inference cost. The combination of a higher-capacity decoder and a lightweight decoder is motivated by their distinct roles across decoding stages.

The evaluation spans multiple code families and reports both error rate and latency, highlighting efficiency as a key contribution.

Critical weaknesses

The experimental comparison does not provide a sufficiently complete basis for assessing the paper’s novelty and performance claims. Several closely related recent methods are either missing from the experiments or not clearly identified. The paper cites the Syndrome-Flow Consistency Model (ECCFM), but it does not appear among the single-step baselines. SB-ECC (Alon Helvits, Eliya Nachmani, Score Based Error Correcting Code Decoder, ICML 2026), which also addresses diffusion-based channel decoding and efficiency trade-offs, is not discussed or compared. Spectral-Aligned Pruning is cited, but its connection to the paper’s lightweight-model design and pruning choices is not clearly established.

TCD appears to be represented by TECCT or TECCT+ in the multi-step comparison table, but the paper does not clarify this mapping. Results are reported for only a subset of the listed code settings, making it difficult to assess the breadth and comparability of the evaluation.

Because the paper’s main claims involve improved decoding and a better latency–performance trade-off, the missing or unclear comparisons make it difficult to determine how much of the observed gain is due to the proposed training and inference design, the model architecture, or what is genuinely new compared to recent work. The paper would be stronger with a clearer account of baseline selection and comparable results for the most relevant methods, where feasible.

Questions for authors
  1. Could you clarify whether TECCT and TECCT+ in the comparison tables are the TCD baseline, and specify the settings and code cases used for that comparison?
  2. Could you compare against ECCFM, the cited single-step consistency-model decoder, or explain why a direct comparison is not feasible?
  3. Could you discuss SB-ECC and clarify how the proposed approach differs from these diffusion-based channel decoders, particularly regarding the paper’s efficiency and inference claims?
  4. How does the pruning used for FAWN relate to Spectral-Aligned Pruning, and which parts of the paper’s lightweight-model contribution should be considered distinct from existing pruning approaches?
3 comments

We thank the reviewer for the helpful comments. We address the baseline comparisons, the relationship to prior work, and the sources of improvement below.

TCD baseline

Thank you for identifying the ambiguity in our baseline naming. The entries labeled TECCT refer to the TCD baseline. We will clarify the variant names and their source for Tables 2 and 4. Specifically, we use the reported TCD-HYB results from Table II of the TCD paper, with feature dimension d=128d=128 and six layers. We have not located a public implementation, so our comparison is limited to the code/SNR settings reported in that work. In these settings, TCD provides competitive decoding performance, while our heterogeneous decoder achieves better results in most entries.

Comparison with more recent methods

We will include comparisons with ECCFM and SB-ECC in the revised manuscript. The following table compares their reported results with ours on four representative codes. Entries are −ln⁡(BER)-\ln(\mathrm{BER}), where higher values indicate better decoding performance.

CodesSNRECCFMDEER 11-stepSBECCOur Hybrid
BCH 63,364dB5.005.545.746.04
BCH 63,365dB6.897.648.128.44
BCH 63,366dB9.7610.6811.2011.46
Polar 128,644dB8.019.519.0310.75
Polar 128,645dB12.2213.3013.1313.98
Polar 128,646dB16.7116.1016.9417.44
Polar 64,324dB7.557.827.777.91
Polar 64,325dB10.3110.2510.3010.48
Polar 64,326dB13.8013.7713.7813.94
CCSDS 128,644dB7.957.888.608.78
CCSDS 128,645dB12.6812.6213.4313.80
CCSDS 128,646dB17.0117.0819.2018.08

Across these 12 code/SNR settings, one-step DEER outperforms ECCFM in seven, while our heterogeneous decoder outperforms SB-ECC in eleven. SB-ECC achieves the better result on CCSDS(128,64) at 6 dB. We will include the complete comparison in the revised manuscript. Our heterogeneous decoder outperforms SB-ECC in 32 of 39 entries, with an average SNR gain of 0.28 dB across all evaluated 44dB operating points in our experimental setups.

Distinction from SB-ECC

Both methods regress continuous denoising fields and perform probability-flow ODE (PF-ODE) decoding. Our framework differs in the following respects:

  1. We explicitly condition the denoising field on the noise level and determine the initial diffusion time from the channel SNR. SB-ECC instead uses a time-unconditional predictor without SNR side information. It starts from the received observation and applies a prescribed noise-scale schedule without determining the initial diffusion time from the channel SNR.
  2. Following prior work that uses symmetry-preserving decoding, we parameterize the field as vθ(x,t)=sign⁡(x)⊙fθ(∣x∣,s(x),t)v_\theta(x,t)=\operatorname{sign}(x)\odot f_\theta(|x|,s(x),t). This construction explicitly enforces codeword-sign equivariance and permits training with the all-zero codeword. SB-ECC instead directly predicts the field from signed observations. Our construction incorporates the known symmetry into the architecture rather than requiring it to be learned from random-codeword examples.
  3. We introduce lightweight FAWN decoders and allocate different decoder capacities across the denoising trajectory, reducing per-step computational cost in later stages.

Architectural optimization and its relationship to Spectral-Aligned Pruning

We thank the reviewer for pointing out the ambiguity in our use of “pruning.” FAWN uses observed sparsity to motivate FFN-width reduction, rather than applying a pruning-mask transfer procedure to a pretrained decoder. We will replace the relevant uses of “pruning” with “architectural optimization” or “FFN-width reduction” in the revised manuscript.

SAP and AECCT motivate the investigation of redundancy in neural decoders. In FAWN, this motivates FFN-width reduction, alongside computational optimizations to EfficientMPT’s ECC attention. These include reordering aggregation and projection, correcting normalization, and executing existing parity-check connections through a specialized sparse gather-and-accumulate kernel. These optimizations do not use SAP’s spectral cross-code pruning-mask reuse or AECCT’s ternary quantization.

AuthorsOpenReview ↗

Comment continued

Sources of improvement

We distinguish improvements from the training/inference formulation, the lightweight architecture, and heterogeneous inference. First, DEER retains the CrossMPT backbone. The single-step results in Table 5 show SNR gains of approximately 0.1–0.3 dB over CrossMPT, without additional inference iterations. Multi-step DEER provides a further gain of approximately 0.4 dB over single-step DEER, at the cost of additional model evaluations. These gains refer to reductions in the required Eb/N0E_b/N_0 at matched BER.

Second, FAWN targets the computational cost of each evaluation. FFN-width reduction and attention-computation optimizations reduce latency, while the increased head count and FAWN-L8’s additional depth improve the accuracy–cost balance. The latency measurements reflect these combined changes.

Finally, heterogeneous inference assigns DEER to the noisier stages and FAWN to later stages. Comparisons with homogeneous decoding and alternative model-order schedules evaluate the benefit of allocating model capacity across the denoising trajectory.

For efficiency, DEER’s implementation optimizations reduce single-pass latency by approximately 25% relative to CrossMPT, while FAWN’s architectural and computational optimizations reduce latency by approximately one third relative to EfficientMPT. Kernel selection is autotuned between dense and sparse parity-check aggregation. Among the configurations in Table 3, the sparse kernel is selected only for Tanner(310,128), where the longer code makes it advantageous. The short-code speedups therefore arise from the other optimizations, while sparse aggregation provides additional savings for the longer configuration.

Thank you for the updated results and explanations. I have updated my rating.

Comments on the paper

4 comments
Area_Chair_Zb1fOpenReview ↗

AC Comment

Thank you for the responses. In addition to the reviewers' comments, I would like to raise a few points. I understand the remaining time is limited, so new experiments are not strictly required. Answers based on existing results, analytical estimates, or clarifications would be appreciated where possible.

1. Figure quality (TcU4, nZZN). Please upload a revised PDF if possible. It should fix the issues the reviewers raised (e.g., Fig. 2) and improve overall legibility, such as the small fonts in Figs. 1 and 4. The PDF currently on OpenReview still contains the original figures.

2. Need for SNR at decoding time. DDECC could also have set its diffusion time from the SNR. As we understand it, it instead uses the syndrome so that it can run without SNR side information. DDECC, ECCT, CrossMPT, and SB-ECC all decode without the channel SNR. Requiring it is therefore a limitation of the proposed method relative to these decoders. Is the SNR actually necessary at inference? The "No SNR" ablation in Table 6 suggests the network itself does not need it. Is it still needed to set t_init and the step schedule? If so, please state this limitation explicitly in the main text. It would also help to show how much performance degrades when the assumed SNR is wrong.

3. Per-codeword latency. All latencies are reported per batch (batch size 2,048), which mainly reflects throughput. For a decoder, per-codeword latency (batch size 1) seems more appropriate. It would also remove the effect where a single slow codeword delays the whole batch, which is how DDECC's whole-batch stopping rule makes it look slow. Please consider reporting latency at batch size 1 throughout the paper.

4. Separating implementation gains from algorithmic gains (related: TcU4). In your response to pc6d, you note that DEER's ~25% latency reduction over CrossMPT comes from implementation optimizations, even though the two share the same architecture. FAWN versus EfficientMPT raises the same issue. If the work is to be evaluated as an algorithmic contribution rather than an engineering one, the comparison should apply the same optimizations to both (or to neither). Please provide such comparisons, or clarify which of the speedups would also apply to the baselines.

5. FLOPs/MACs. Latency depends heavily on implementation. FLOPs/MACs (together with the number of function evaluations) would be a fairer basis for separating algorithmic improvement from implementation optimization. The paper currently reports only parameter counts. In particular, please compare CrossMPT vs. DEER and EfficientMPT vs. FAWN in the single-step setting under matched FLOPs and latency. The backbone should be kept the same, so that the effect of the proposed formulation can be seen on its own.

6. Cost comparison with SB-ECC (pc6d). Thank you for the new BER/BLER comparisons with SB-ECC and ECCFM. However, the paper's main claim concerns the BER–latency trade-off. Please also provide latency and/or FLOPs (and NFE) for these methods alongside the multi-step Hybrid.

AuthorsOpenReview ↗

Response to Area Chair Zb1f

We thank the Area Chair for these constructive comments. Below we clarify the inference assumptions, distinguish algorithmic and implementation gains, and provide additional measurements at 4 dB. Further details and results shall be included in a revised manuscript.

Figure quality

We have uploaded a revised manuscript correcting Fig. 2, aligning the diffusion convention in Fig. 1, and providing additional experimental results as we stated in prior responses to reviewers.

We shall address the new results and clarification in response to this comment in a future revision.

The Need for SNR at Inference

Yes, the current multi-step inference procedure requires a channel-SNR value to determine tinitt_\text{init} and the time schedule. The "No SNR" ablation removes the network's SNR embedding, but not the scheduling requirement. We will state it explicitly in the main text.

Accurate non-data-aided SNR estimation for BPSK over AWGN is well established, including moment-based and maximum-likelihood approaches [1,2]. Its accuracy depends on the observation length and channel conditions; averaging across codewords provides additional samples when the channel remains stable.

DEER’s existing results also involve a small training–inference noise-level mismatch: auxiliary Gaussian perturbations are included during training but omitted at inference. Its observed performance therefore demonstrates tolerance to this specific mismatch, although this does not replace an evaluation of sensitivity to SNR-estimation errors.

  1. A. Wiesel, J. Goldberg and H. Messer, "Non-data-aided signal-to-noise-ratio estimation," 2002 IEEE International Conference on Communications. Conference Proceedings. ICC 2002 (Cat. No.02CH37333), New York, NY, USA, 2002, pp. 197-201 vol.1, doi: 10.1109/ICC.2002.996844.
  2. D. R. Pauluzzi and N. C. Beaulieu, "A comparison of SNR estimation techniques for the AWGN channel," in IEEE Transactions on Communications, vol. 48, no. 10, pp. 1681-1691, Oct. 2000, doi: 10.1109/26.871393.

Per-codeword Latency

Our original evaluation targeted throughput-oriented GPU decoding, where processing independent codewords in parallel can exploit a practical strength of neural decoders. Batched measurements are therefore relevant to this throughput-oriented regimes, although they do not characterize isolated-codeword latency or batch-formation delay. We now additionally report mean and 90th-percentile (P90) decoding latency at batch size one and 4 dB. This also removes the whole-batch stopping effect in the DDECC comparison.

Here we provide the batch 1 decoding latency comparison at 44 dB. All methods below were evaluated on an NVIDIA RTX 4090 using the manuscript’s timing protocol, with batch size one. We also use early stopping; its implementation details will be included in the main text. Across the four evaluated 4 dB settings, we observed no aggregate BER degradation and a 0.47% relative reduction in aggregate BLER.

We will report the early-exit protocol explicitly and distinguish its results from the fixed-step evaluation. The fixed-step comparisons assess heterogeneous allocation, whereas the present measurements characterize per-codeword decoding with early stopping. Early stopping terminates decoding when the estimated codeword satisfies all parity checks. P90 is the 90th percentile of individual-codeword decoding latency.

CodeHybrid Mean/P90 LatencySB-ECC Mean/P90 LatencyDDECC( Mean with LS)DDECC( Mean without LS)
BCH (63,36)1.55 ms/1.17ms2.51 ms/4.00 ms2.81 ms10.53 ms
Polar (64,32)1.13 ms/1.12ms2.74 ms/3.92 ms1.26 ms11.26 ms
Polar (128,64)1.10 ms/1.12ms3.19 ms/3.97ms--
CCSDS (128,64)1.14 ms/1.10ms3.20 ms/3.21ms--

DDECC measurements are limited to two codes, as we have not had sufficient time to complete the remaining evaluations. Our decoder achieves lower measured latency and mean NFE than SB-ECC under the evaluated configurations; the latency difference also includes implementation effects.

To reduce avoidable implementation overhead, we restricted DDECC’s CPU numerical libraries to one thread, reducing overhead from small CPU operations. For SB-ECC, we implemented a batch-size-one specialization that removes redundant active-word gathers, clones, and scatter updates. On the checked inputs, this specialization preserved decoding decisions and convergence steps. The table reports these optimized baseline implementations.

These changes address identifiable baseline overheads and strengthen the runtime comparison, although they do not establish identical implementation efficiency across methods.

AuthorsOpenReview ↗

Response to Area Chair Zb1f Part 2

Gain Separation

Our formulation contribution concerns the training target and inference dynamics. DEER retains the CrossMPT backbone, allowing accuracy differences to be examined without increasing backbone capacity or the number of evaluations in the single-step setting. The approximately 25% single-pass latency reduction is attributable to implementation optimizations and should not be credited to the formulation.

FAWN combines FFN-width reduction and modified head/depth configurations with computation reordering, normalization changes, and sparse parity-check aggregation. Some implementation improvements could also benefit EfficientMPT. We will distinguish these contributions and clarify that the reported latency reduction reflects their combined effect in the following part.

The current FAWN results combine formulation, architectural, and implementation changes rather than fully isolating each factor. We therefore restrict our attribution accordingly: the unchanged-backbone single-step DEER comparison supports the formulation’s accuracy benefit, while the FAWN ablations characterize the combined efficiency changes.

FLOPs/MACs

We have been showing the real-world latency data instead of measured or estimated number of operations since our implementation uses auto-tuned kernels where the actual computation differ for different linear codes, and the intrinsic operation counting library does not support our implemented kernels. We now provide profiler-based FLOP estimates using the counting procedure described below. We will document the included operations and treatment of custom kernels in the revised appendix.

CodeDEER FLOPs/callFAWN L8 FLOPs/callFAWN L6 FLOPs/callMean Hybrid FLOPsEfficientMPT FLOPs/callSB-ECC FLOPs/callMean SB-ECC FLOPs
BCH (63,36)222.96M149.50M112.24M0.294G200.326M222.83M0.646G
Polar (64,32)239.25M160.45M120.48M0.240G214.458M239.12M0.821G
Polar (128,64)503.58M338.21M254.20M0.504G442.565 M503.45M1.947G
CCSDS (128,64)503.58M338.21M254.20M0.516G442.565M503.45M1.963G

We obtain FLOP counts using PyTorch’s FlopCounterMode, counting one multiplication and one addition as two FLOPs. For profiling, FlexAttention is replaced by mathematically equivalent dense attention. The reported values therefore represent dense-equivalent model FLOPs rather than the operations executed by the optimized kernels.

We next examine successive computation and architectural changes to EfficientMPT using single-pass latency. PCM projection refers to the projection kernel in our implementation, while reorder refers to our reordering of computation specified in Appendix C. FAWN model additionally changes the FFN size along with modified normalization for training stability. The following tables report single-pass execution times at batch sizes 1 and 2048, measured on an NVIDIA RTX 4090 using the manuscript’s timing protocol.

Batch size 1

CodeEfficientMPT+ PCM projection+ PCM and reorderedFAWN L6 H8
BCH (63,36)0.490 ms0.479 ms0.488 ms0.495 ms
Polar (64,32)0.483 ms0.498 ms0.504 ms0.524 ms
Polar (128,64)0.489 ms0.503 ms0.508 ms0.515 ms
CCSDS (128,64)0.518 ms0.524 ms0.526 ms0.559 ms

Batch size 2,048

CodeEfficientMPT+ PCM projection+ PCM and reorderedFAWN L6 H8
BCH (63,36)25.126 ms23.424 ms23.264 ms15.340 ms
Polar (64,32)26.273 ms24.404 ms24.579 ms16.890 ms
Polar (128,64)57.477 ms51.929 ms51.803 ms35.357 ms
CCSDS (128,64)57.578 ms52.130 ms52.256 ms37.217 ms

At batch size one, the modifications provide no consistent single-pass latency reduction; FAWN-L6-H8 is slightly slower in these settings. At batch size 2048, it reduces batch execution time by approximately 35–39% relative to EfficientMPT. The intermediate columns show smaller savings from transferable computation optimizations. H8 is used here to control head count for fair comparison; our main FAWN models use H16.

AuthorsOpenReview ↗

Response to Area Chair Zb1f Part 3

The corresponding mean numbers of function evaluations under early stopping are reported below.

CodeHybridSB-ECC
BCH (63,36)1.462.90
Polar (64,32)1.003.43
Polar (128,64)1.003.87
CCSDS (128,64)1.033.90

The low mean NFE suggests that many frames terminate during the initial DEER stage; these measurements therefore primarily characterize early-exit efficiency, while the fixed-step comparisons assess heterogeneous allocation.

Cost comparison with SB-ECC

At 4 dB, the evaluated early-exit decoder uses fewer mean function evaluations and lower estimated total FLOPs than SB-ECC on all four tested codes. It also achieves lower measured batch-size-one latency under the implementations evaluated here. These timing differences include implementation effects.

SB-ECC uses CrossMPT as its backbone. For the matched configuration evaluated here, its per-call FLOPs provide a reference for CrossMPT’s backbone computation, subject to additional SB-ECC-specific operations.

Given ECCFM’s comparable decoding performance to single-step DEER, we use single-step DEER rather than the multi-step Hybrid for comparison. For ECCFM’s matched CrossMPT configuration, the SB-ECC per-call FLOPs reported above provide an approximate reference for its single-pass backbone computation, allowing a cost comparison at similar decoding performance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.