{"id":"ce18bf00-b9ff-4b27-8ebc-acede6849da3","arxiv_id":"2505.17834","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A hybrid Mamba-Transformer decoder with parity-check-aware masking and progressive supervision outperforms prior neural decoders on multiple error-correcting codes.","lead":"The paper introduces a hybrid neural decoder that mixes Mamba and Transformer layers to decode error-correcting codes, reporting faster and more accurate decoding than prior learned methods on several code families. A generalist might care because better and faster decoding could improve real-time communication systems and storage.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 13-14 index f(H) as f(H)[l,d] with l up to L and d up to D, but f(H) is (n-k)x(2n-k); the code-aware Mamba mask is undefined as written and the central mechanism is not reproducible.","rationale":"Read in good faith, the paper has real strengths: it reports a broad benchmark table, an ablation table isolating the three proposed components, and a plausible efficiency argument from Mamba's linear state-space complexity. The central claim, however, is specifically that the hybrid decoder with code-aware masking outperforms Transformer-only and standard Mamba decoders. For that claim to be true, the code-aware masking must actually be the mechanism described. The equations in Section 4.0.1 contain a dimension mismatch that is not a minor typo: it changes which entries of \\bar B and C are zeroed. If the intended transpose is supplied, the mask may indeed encode parity-check connectivity; if not, the reported f(H) benefit could be an artifact of an unstated reshaping. This is exactly the kind of internal inconsistency that conditional acceptance should require resolving. A missing pure-Mamba baseline is also a concern for the 'standard Mamba models' part of the abstract, but the mask mismatch is more load-bearing because it affects the validity of the core mechanism. I do not see evidence of bad faith; the issue is incomplete and unclear specification. The correct disposition is to keep the reader's CONDITIONAL verdict pending a corrected derivation or released code.","tokens_in":10298,"tokens_out":7297,"duration_ms":77784,"concrete_test":"Run the BCH(63,45) experiments under a literal correction: replace every f(H)[l,d] in Eqs. 13-14 with f(H)^T[l,d] (equivalently, define the mask as [H; I_{n-k}]^T), keeping everything else identical, and compare the full method against ablation (i) in Table 2. If the corrected implementation does not reproduce the reported -ln(BER) values (7.01/10.12/14.26 for full, 6.86/9.88/13.76 for (i)) within Monte Carlo error, or if a literal implementation of Eq. 13 fails due to index bounds, then the claimed benefit of the proposed mask is not supported as described.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that Eqs. 13-14 implement a code-aware mask on the Mamba state updates. As written, they do not define a valid operation. f(H) is defined in Eq. 4 as [H; I_{n-k}], a matrix of shape (n-k) x (2n-k), with row index over parity checks and column index over sequence positions. Eq. 13 writes \\bar B_i^M[l,d,s] = f(H)[l,d] \\bar B_i[l,d,s] for l in [1,L], d in [1,D], and Eq. 14 similarly uses f(H)[l,s]. Since L=2n-k and D=128 while f(H) has only n-k rows, f(H)[l,d] is out of bounds whenever l>n-k and d>2n-k. The only self-consistent reading is that the mask is applied transposed, f(H)^T, but the transposition is never stated, and the condition d<(n-k) in Eq. 13 does not match the stated row/column semantics. Moreover, the mask is applied only to \\bar B and C, not to the state-transition matrix \\bar A, so the claimed 'parity-check-line restriction' is not guaranteed by the displayed equations. Because the central contribution and the ablation improvement in Table 2 hinge on f(H), the mechanism is underspecified to the point of being non-reproducible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ECCM, a hybrid decoder for binary linear block codes that alternates Mamba and Transformer layers. It introduces a parity-check-derived mask f(H) intended to restrict Mamba state updates to parity-check-line interactions, and a progressive layer-wise BCE loss that supervises intermediate outputs. Experiments on BCH, Polar, LDPC, and MacKay codes compare against BP, ARBP, ECCT, AECCT, and CrossMPT, reporting gains in negative log-BER and in inference speed, with an ablation on BCH(63,45).","tokens_in":10652,"tokens_out":4680,"duration_ms":48789,"significance":"If the architecture and masking mechanism were implemented as described and the results held with proper statistical support, the work would offer a useful efficiency-accuracy trade-off for neural decoding. The systematic multi-code evaluation, early-stopping speed measurements, and inclusion of MacKay codes are strengths. However, the current significance is limited by the undefined f(H) mask procedure, the absence of the claimed Mamba-only baseline, and the lack of code or error bars. The paper does not ship code or machine-checked proofs, so reproducibility rests entirely on the written specification, which is currently incomplete.","major_comments":[{"comment":"The code-aware mask f(H) is defined in Eq. (4) as [H; I_{n-k}], so f(H) has (n-k) rows and 2n-k columns. Equation (13) indexes f(H)[l,d] with l in [1,L] and d in [1,D], where L=2n-k and D=128, and Eq. (14) indexes f(H)[l,s] with s in [1,S]. These indices exceed the dimensions of f(H) whenever l>n-k, d>2n-k, or s>2n-k. The condition d < (n-k) in Eq. (13) does not fix the problem because the first index l still ranges to L. As written, the mask cannot be evaluated, so the central mechanism and the ablation in Table 2 are not reproducible. The authors must specify the intended orientation (e.g., f(H)^T), define how the D and S axes align with the parity-check structure, and provide the exact indexing, preferably with released code.","section":"4.0.1, Eqs. (13)-(14)"},{"comment":"The abstract and conclusion claim that ECCM \"significantly outperforms ... standard Mamba models,\" but no Mamba-only baseline appears in Table 1 or in the ablation of Table 2; the only non-hybrid variant in Table 2 is Transformer-only. A decoder with only Mamba layers, with and without the proposed mask, must be evaluated before this claim can be made.","section":"Abstract and Section 6, Table 1"},{"comment":"All reported results are point estimates. No error bars, number of random seeds, or statistical significance tests are provided, so statements such as \"significant improvement\" and \"over 18%\" are not substantiated. The paper should report variance across at least several training runs for the proposed method and ideally for baselines, and should release code to allow independent verification.","section":"Section 6, Tables 1-2"},{"comment":"Even after repairing the indexing, the mask is applied only to B_i^M and C_i^M, not to the state-transition matrix A-bar in the recurrence h_l[d,s] = A-bar[l,d,s] h_{l-1}[d,s] + B_i^M[l,d,s] u_i_conv[l,d]. Because A-bar is unmasked, information from earlier positions propagates through the state regardless of the B mask. The paper's claim that the f(H) mask \"ensures that interactions only happens along the parity check lines\" (Section 4) is therefore not entailed by the displayed equations; the authors should specify how the state-transition path is restricted or soften the claim.","section":"4.0.1, Eq. (15)"}],"minor_comments":[{"comment":"The caption reads \"EECM architecture\" but the model is called ECCM throughout the paper.","section":"Figure 1 caption"},{"comment":"\"hattention heads\" should read \"h attention heads,\" and the reshaping notation should specify the tensor permutation explicitly.","section":"Eq. (22)"},{"comment":"Equation (30) indexes y_in[l] for the estimated codeword, but y_in has length L=2n-k while the codeword estimate should have length n; the indexing should clarify which components of y_in are used.","section":"Eq. (30)"},{"comment":"The state h_l is written as h_l[d,s] while A-bar includes an explicit sequence index l, so the state should be denoted h[l,d,s] or the notation made consistent to avoid ambiguity.","section":"Eq. (15)"},{"comment":"The sentence \"We use the zero-codeword in the training process ... by simply changing to random codewords on model evaluation\" is confusing and should be rewritten to state explicitly that training uses the zero codeword and evaluation uses random codewords.","section":"Section 5"},{"comment":"The text says \"Examining the attention maps Fig 5,\" but the attention maps are presented as Table 5; the cross-reference should be corrected.","section":"Section 7.3"},{"comment":"The complexity expression in Section 7.2 has an unmatched parenthesis and should be rewritten with balanced parentheses and consistent notation.","section":"Section 7.2"}],"recommendation":"major_revision","confidential_remarks":"The paper builds heavily on the authors' prior ECCT/AECCT line and on CrossMPT, so the novelty is incremental. The central problem is that the f(H) mask, which is the main architectural contribution, is undefined as written; this is fixable but requires a careful restatement and ideally code. I also recommend requiring the Mamba-only baseline before any claims about outperforming standard Mamba models are made. The missing error bars are a concern for a journal submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about this paper before spending time on it. First, it's a plausible step forward for learned decoders: combining Mamba layers with transformer attention, plus a code-aware mask and per-layer loss, gives real BER gains and 3-4x faster inference on several short BCH and Polar codes compared to ECCT/AECCT/CrossMPT. The experiments are broad and the ablation is useful. Second, the equations that define the new f(H) mask do not parse. Eq. 13 indexes f(H)[l,d] for l=1..L and d=1..D, but f(H) is an (n-k) x (2n-k) matrix. The only self-consistent reading is that they meant f(H)^T, and even then it's unclear why masking B and C (not A) enforces the claimed parity-check-line restriction on state transitions. Since the ablation credits this mask with a nontrivial chunk of the gain, the central mechanism is non-reproducible as written. That is a load-bearing flaw, not a typo.\n\nWhat's genuinely new: the hybrid Mamba-Transformer decoder itself, and the layer-wise loss applied at each decoding stage. The speed measurements are concrete and reproducible in principle. The citation pattern is honest; the authors are building on their own ECCT/AECCT work and cite CrossMPT properly.\n\nSoft spots in proportion: (1) The abstract claims superiority over 'standard Mamba models,' but there is no Mamba-only baseline in any table. The ablation compares hybrid vs transformer-only, not vs Mamba-only. (2) No error bars, no code, no statistical significance tests. This is unfortunately common in this subfield, but it limits the strength of the claims, especially when differences are ~4%. (3) The mask equations are the real problem. If the authors can provide a clean statement, e.g., define f(H)^T and show how it restricts which positions can affect a state, the paper could be salvageable with moderate work.\n\nBottom line: this is a serious paper in intent and scope, and the empirical direction is worth exploring. A good referee should not desk-reject it, but should send it back with major revisions. The mask definition must be fixed and a Mamba-only baseline added before the claimed gains can be trusted. If I were the editor, I'd send it to review but with a strong note that the central mechanism is currently underspecified. I would not cite it myself until the equations are corrected.\n\nTake it to a reading group? Maybe, if you want a case study in how a promising architecture can be undermined by a notation error.","headline":"Promising hybrid decoder with real empirical gains, but the defining f(H) mask equations are dimensionally invalid and the central mechanism is not reproducible as written.","tokens_in":11075,"tokens_out":3165,"would_cite":false,"duration_ms":32985,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid decoder that alternates Mamba layers, an efficient state-space sequence model, with Transformer attention layers, masks both with the code's parity-check structure, and supervises every layer, outperforming Transformer-only and…","keywords":["learned decoding","error-correcting codes","Mamba state-space model","Transformer attention","parity-check masking","progressive layer-wise loss","BCH and Polar codes","LDPC codes"],"falsifier":"Inspect the forward-pass tensor shapes for BCH(63,45), where $n-k=18$ and $L=81$: Eq. 13 requires $f(H)[l,d]$ for $l$ up to 81 even though $f(H)$ is declared to have only 18 rows; if no transposition or padding rescues the formula, then re-running the Table 2 ablation with the mask corrected, expected to reproduce $-\\ln(\\mathrm{BER})$ values 7.01/10.12/14.26 for the full model and 6.86/9.88/13.76 with the $g(H)$ mask, would settle whether the mask is responsible for the gain.","tokens_in":10140,"feed_emoji":"📡","tokens_out":16079,"duration_ms":111886,"temperature":0.7,"pith_summary":"The paper tries to establish that a decoder alternating Mamba layers, an efficient state-space sequence model, with Transformer attention layers can beat both a Transformer-only decoder and a standard Mamba decoder for binary linear block codes. The proposed design, called ECCM, restricts the Mamba layers' state updates with a mask built from the code's parity-check matrix, and adds a progressive loss that supervises the output of every layer. The reported experiments show consistent gains across BCH, Polar, and MacKay codes, up to an 18 percent relative improvement in $-\\ln(\\mathrm{BER})$ on some BCH and Polar codes, roughly comparable results on LDPC codes, and lower measured inference latency than the compared Transformer decoders. If the claim holds, learned decoding becomes simultaneously more accurate and faster, which matters for real-time communication and storage systems.","feed_headline":"Hybrid Mamba-Transformer decoder beats single-architecture decoders","feed_subtitle":"Parity-check masking and layer-wise loss lift BCH and Polar decoding accuracy by up to 18 percent.","key_machinery":"The central object is the mask $f(H) = [H; I_{n-k}]$, a 0/1 matrix built from the parity-check matrix $H$ and an identity block; applied inside each Mamba block to the discretized state-space matrices, it is intended to make a position's state update depend only on positions that share a parity check. The Transformer blocks use the earlier $g(H)$ mask from ECCT, which restricts attention to pairs lying on a common parity-check line, giving the hybrid a global but code-structured receptive field. The third mechanism is a progressive layer-wise loss: each layer's output is projected to an error estimate, the binary cross-entropy with the true error vector is summed over layers, and inference stops early once the syndrome of the current estimate matches the received syndrome. Together these parts are meant to let the model combine Mamba's linear-cost sequential processing with Transformer's global context while keeping the code structure explicit at every depth.","core_discovery":"On its own terms, the paper's discovery is that a hybrid decoder is better than either of its ingredients: interleaving Mamba blocks and Transformer blocks, with a parity-check-derived mask applied inside each Mamba block and a binary cross-entropy loss summed over every layer, yields lower bit error rates than the ECCT family, the accelerated HPSA variant, and CrossMPT on most tested codes. The reported measurements show ECCM surpassing CrossMPT by a double-digit relative margin in $-\\ln(\\mathrm{BER})$ on BCH(63,45), and improving on Polar(128,86) by about 7 percent at the lowest tested SNR, while trading small differences on LDPC codes and doing so at per-codeword inference times roughly two to four times faster in the measured runs. The paper attributes these gains jointly to the hybrid layout, the new $f(H)$ mask, and the layer-wise loss, with ablations on BCH(63,45) showing that removing any one of the three reduces performance.","pith_inferences":["The paper does not test longer codes; because the Mamba path has linear cost in sequence length, the speed advantage over attention-only decoders should widen as block length grows beyond $n=128$, which is a testable scaling prediction.","If the apparent dimension mismatch in Eq. 13 is real, the credit assigned to the $f(H)$ mask is unverified; a corrected re-run of the BCH(63,45) ablation would either confirm the masking mechanism or redistribute the gain to the hybrid layout and the layer-wise loss.","The syndrome-based early stopping is a general compute-adaptive idea; it could be ported to belief-propagation-style neural decoders or diffusion decoders to give them the same latency flexibility, something the paper does not explore.","The design principle, structured sparsity in the sequential path and structured attention in the global path, could generalize beyond channel decoding to any sequence task with known constraint structure, though the paper does not claim this."],"forward_implications":["On the tested BCH and Polar codes, ECCM's $-\\ln(\\mathrm{BER})$ exceeds the best previous neural decoder by up to 18 percent.","On LDPC(121,60), ECCM trails CrossMPT by about 0.39 at 5 dB and 0.55 at 6 dB while running faster, and the paper notes that combining ECCM with CrossMPT may close that gap.","The early-stopping rule ties computation to channel quality: at 6 dB the average inference time drops to about 55 percent of the no-early-stopping time on LDPC(121,60) and Polar(128,96), since later layers are skipped.","The ablations attribute each gain to a separate design choice, so the three components, hybrid layout, $f(H)$ mask, and multi-layer loss, each earn their place.","The method carries over to any linear block code with a known parity-check matrix, since both masks are constructed directly from $H$.","The syndrome-based early-stopping rule means that the same trained model adapts its latency to channel quality without retraining."],"supporting_citations":[{"why":"Supplies the Mamba selective state-space model that forms the sequential half of the hybrid decoder.","marker":"[17]"},{"why":"Defines the ECCT Transformer decoder, the g(H) parity-check mask, and the syndrome-based input notation that ECCM extends.","marker":"[8]"},{"why":"Provides the HPSA attention mechanism used in the Transformer blocks and the AECCT baseline.","marker":"[27]"},{"why":"CrossMPT is the main state-of-the-art learned decoder against which ECCM is measured on accuracy and inference speed.","marker":"[34]"},{"why":"Introduces the syndrome-plus-magnitude input and zero-codeword training procedure used in all experiments.","marker":"[3]"},{"why":"Supplies the Transformer architecture whose global self-attention is the other half of the hybrid.","marker":"[35]"},{"why":"Provides the parity-check matrices for the BCH, LDPC, Polar, and MacKay codes used in training and evaluation.","marker":"[21]"}],"fun_headline_variants":["Hybrid Mamba-Transformer decoder lifts BCH/Polar accuracy by up to 18%","Masked Mamba layers plus Transformer blocks cut decoding errors","Interleaved Mamba and Transformer decode faster and better than either","Layer-wise loss and parity mask boost hybrid decoder performance","Hybrid decoder: Mamba efficiency, Transformer reach, 18% gain on codes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the $f(H)$ mask is applied to the Mamba state updates exactly as the equations intend; the indexing in Eq. 13 does not match the mask's declared dimensions, so the parity-check restriction that the reported gains depend on is not established from the text alone.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid Mamba-Transformer decoder lifts BCH/Polar accuracy by up to 18%","Masked Mamba layers plus Transformer blocks cut decoding errors","Interleaved Mamba and Transformer decode faster and better than either","Layer-wise loss and parity mask boost hybrid decoder performance","Hybrid decoder: Mamba efficiency, Transformer reach, 18% gain on codes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00021,"raw_usage":{"total_tokens":1367,"prompt_tokens":858,"completion_tokens":509,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":413}},"tokens_in":474,"tokens_out":509,"duration_ms":5016,"temperature":1.0,"reasoning_tokens":413,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:39:03.235951+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the forward-pass tensor shapes for BCH(63,45), where $n-k=18$ and $L=81$: Eq. 13 requires $f(H)[l,d]$ for $l$ up to 81 even though $f(H)$ is declared to have only 18 rows; if no transposition or padding rescues the formula, then re-running the Table 2 ablation with the mask corrected, expected to reproduce $-\\ln(\\mathrm{BER})$ values 7.01/10.12/14.26 for the full model and 6.86/9.88/13.76 with the $g(H)$ mask, would settle whether the mask is responsible for the gain.","supporting_citations":[{"cited_title":"Error correction code transformer, 2022","cited_arxiv_id":null,"evidence_quote":"Defines the ECCT Transformer decoder, the g(H) parity-check mask, and the syndrome-based input notation that ECCM extends."},{"cited_title":"Accelerating error correction code transformers, 2025","cited_arxiv_id":null,"evidence_quote":"Provides the HPSA attention mechanism used in the Transformer blocks and the AECCT baseline."},{"cited_title":"Deep learning for decoding of linear codes-a syndrome-based approach","cited_arxiv_id":null,"evidence_quote":"Introduces the syndrome-plus-magnitude input and zero-codeword training procedure used in all experiments."},{"cited_title":"Database of channel codes and ml simulation results","cited_arxiv_id":null,"evidence_quote":"Provides the parity-check matrices for the BCH, LDPC, Polar, and MacKay codes used in training and evaluation."}],"review_version":1}